ArmenianChinese (Traditional)EnglishGermanRussian
The Python Language Reference -

The Python Language Reference

Language Reference

This reference manual describes the syntax and “core semantics” of the language. It is terse, but attempts to be exact and complete. The semantics of non-essential built-in object types and of the built-in functions and modules are described in The Python Standard Library. For an informal introduction to the language, see The Python Tutorial. For C or C++ programmers, two additional manuals exist: Extending and Embedding the Python Interpreter describes the high-level picture of how to write a Python extension module, and the Python/C API Reference Manual describes the interfaces available to C/C++ programmers in detail.

1. Introduction

This reference manual describes the Python programming language. It is not intended as a tutorial.

While I am trying to be as precise as possible, I chose to use English rather than formal specifications for everything except syntax and lexical analysis. This should make the document more understandable to the average reader, but will leave room for ambiguities. Consequently, if you were coming from Mars and tried to re-implement Python from this document alone, you might have to guess things and in fact you would probably end up implementing quite a different language. On the other hand, if you are using Python and wonder what the precise rules about a particular area of the language are, you should definitely be able to find them here. If you would like to see a more formal definition of the language, maybe you could volunteer your time — or invent a cloning machine :-).

It is dangerous to add too many implementation details to a language reference document — the implementation may change, and other implementations of the same language may work differently. On the other hand, CPython is the one Python implementation in widespread use (although alternate implementations continue to gain support), and its particular quirks are sometimes worth being mentioned, especially where the implementation imposes additional limitations. Therefore, you’ll find short “implementation notes” sprinkled throughout the text.

Every Python implementation comes with a number of built-in and standard modules. These are documented in The Python Standard Library. A few built-in modules are mentioned when they interact in a significant way with the language definition.

1.1. Alternate Implementations

Though there is one Python implementation which is by far the most popular, there are some alternate implementations which are of particular interest to different audiences.

Known implementations include:

CPython

This is the original and most-maintained implementation of Python, written in C. New language features generally appear here first.

Jython

Python implemented in Java. This implementation can be used as a scripting language for Java applications, or can be used to create applications using the Java class libraries. It is also often used to create tests for Java libraries. More information can be found at the Jython website.

Python for .NET

This implementation actually uses the CPython implementation, but is a managed .NET application and makes .NET libraries available. It was created by Brian Lloyd. For more information, see the Python for .NET home page.

IronPython

An alternate Python for .NET. Unlike Python.NET, this is a complete Python implementation that generates IL, and compiles Python code directly to .NET assemblies. It was created by Jim Hugunin, the original creator of Jython. For more information, see the IronPython website.

PyPy

An implementation of Python written completely in Python. It supports several advanced features not found in other implementations like stackless support and a Just in Time compiler. One of the goals of the project is to encourage experimentation with the language itself by making it easier to modify the interpreter (since it is written in Python). Additional information is available on the PyPy project’s home page.

Each of these implementations varies in some way from the language as documented in this manual, or introduces specific information beyond what’s covered in the standard Python documentation. Please refer to the implementation-specific documentation to determine what else you need to know about the specific implementation you’re using.

1.2. Notation

The descriptions of lexical analysis and syntax use a modified BNF grammar notation. This uses the following style of definition:

name      ::=  lc_letter (lc_letter | "_")* lc_letter ::= "a"..."z"

The first line says that a name is an lc_letter followed by a sequence of zero or more lc_letters and underscores. An lc_letter in turn is any of the single characters 'a' through 'z'. (This rule is actually adhered to for the names defined in lexical and grammar rules in this document.)

Each rule begins with a name (which is the name defined by the rule) and ::=. A vertical bar (|) is used to separate alternatives; it is the least binding operator in this notation. A star (*) means zero or more repetitions of the preceding item; likewise, a plus (+) means one or more repetitions, and a phrase enclosed in square brackets ([ ]) means zero or one occurrences (in other words, the enclosed phrase is optional). The * and + operators bind as tightly as possible; parentheses are used for grouping. Literal strings are enclosed in quotes. White space is only meaningful to separate tokens. Rules are normally contained on a single line; rules with many alternatives may be formatted alternatively with each line after the first beginning with a vertical bar.

In lexical definitions (as the example above), two more conventions are used: Two literal characters separated by three dots mean a choice of any single character in the given (inclusive) range of ASCII characters. A phrase between angular brackets (<...>) gives an informal description of the symbol defined; e.g., this could be used to describe the notion of ‘control character’ if needed.

Even though the notation used is almost the same, there is a big difference between the meaning of lexical and syntactic definitions: a lexical definition operates on the individual characters of the input source, while a syntax definition operates on the stream of tokens generated by the lexical analysis. All uses of BNF in the next chapter (“Lexical Analysis”) are lexical definitions; uses in subsequent chapters are syntactic definitions.

2. Lexical analysis

A Python program is read by a parser. Input to the parser is a stream of tokens, generated by the lexical analyzer. This chapter describes how the lexical analyzer breaks a file into tokens.

Python reads program text as Unicode code points; the encoding of a source file can be given by an encoding declaration and defaults to UTF-8, see PEP 3120 for details. If the source file cannot be decoded, a SyntaxError is raised.

2.1. Line structure

A Python program is divided into a number of logical lines.

2.1.1. Logical lines

The end of a logical line is represented by the token NEWLINE. Statements cannot cross logical line boundaries except where NEWLINE is allowed by the syntax (e.g., between statements in compound statements). A logical line is constructed from one or more physical lines by following the explicit or implicit line joining rules.

2.1.2. Physical lines

A physical line is a sequence of characters terminated by an end-of-line sequence. In source files and strings, any of the standard platform line termination sequences can be used – the Unix form using ASCII LF (linefeed), the Windows form using the ASCII sequence CR LF (return followed by linefeed), or the old Macintosh form using the ASCII CR (return) character. All of these forms can be used equally, regardless of platform. The end of input also serves as an implicit terminator for the final physical line.

When embedding Python, source code strings should be passed to Python APIs using the standard C conventions for newline characters (the \n character, representing ASCII LF, is the line terminator).

2.1.3. Comments

A comment starts with a hash character (#) that is not part of a string literal, and ends at the end of the physical line. A comment signifies the end of the logical line unless the implicit line joining rules are invoked. Comments are ignored by the syntax.

2.1.4. Encoding declarations

If a comment in the first or second line of the Python script matches the regular expression coding[=:]\s*([-\w.]+), this comment is processed as an encoding declaration; the first group of this expression names the encoding of the source code file. The encoding declaration must appear on a line of its own. If it is the second line, the first line must also be a comment-only line. The recommended forms of an encoding expression are

# -*- coding:  -*-

which is recognized also by GNU Emacs, and

# vim:fileencoding=

which is recognized by Bram Moolenaar’s VIM.

If no encoding declaration is found, the default encoding is UTF-8. In addition, if the first bytes of the file are the UTF-8 byte-order mark (b'\xef\xbb\xbf'), the declared file encoding is UTF-8 (this is supported, among others, by Microsoft’s notepad).

If an encoding is declared, the encoding name must be recognized by Python (see Standard Encodings). The encoding is used for all lexical analysis, including string literals, comments and identifiers.

2.1.5. Explicit line joining

Two or more physical lines may be joined into logical lines using backslash characters (\), as follows: when a physical line ends in a backslash that is not part of a string literal or comment, it is joined with the following forming a single logical line, deleting the backslash and the following end-of-line character. For example:

if 1900 < year < 2100 and 1 <= month <= 12 \
   and 1 <= day <= 31 and 0 <= hour < 24 \
   and 0 <= minute < 60 and 0 <= second < 60:   # Looks like a valid date
        return 1

A line ending in a backslash cannot carry a comment. A backslash does not continue a comment. A backslash does not continue a token except for string literals (i.e., tokens other than string literals cannot be split across physical lines using a backslash). A backslash is illegal elsewhere on a line outside a string literal.

2.1.6. Implicit line joining

Expressions in parentheses, square brackets or curly braces can be split over more than one physical line without using backslashes. For example:

month_names = ['Januari', 'Februari', 'Maart',      # These are the
               'April',   'Mei',      'Juni',       # Dutch names
               'Juli',    'Augustus', 'September',  # for the months
               'Oktober', 'November', 'December']   # of the year

Implicitly continued lines can carry comments. The indentation of the continuation lines is not important. Blank continuation lines are allowed. There is no NEWLINE token between implicit continuation lines. Implicitly continued lines can also occur within triple-quoted strings (see below); in that case they cannot carry comments.

2.1.7. Blank lines

A logical line that contains only spaces, tabs, formfeeds and possibly a comment, is ignored (i.e., no NEWLINE token is generated). During interactive input of statements, handling of a blank line may differ depending on the implementation of the read-eval-print loop. In the standard interactive interpreter, an entirely blank logical line (i.e. one containing not even whitespace or a comment) terminates a multi-line statement.

2.1.8. Indentation

Leading whitespace (spaces and tabs) at the beginning of a logical line is used to compute the indentation level of the line, which in turn is used to determine the grouping of statements.

Tabs are replaced (from left to right) by one to eight spaces such that the total number of characters up to and including the replacement is a multiple of eight (this is intended to be the same rule as used by Unix). The total number of spaces preceding the first non-blank character then determines the line’s indentation. Indentation cannot be split over multiple physical lines using backslashes; the whitespace up to the first backslash determines the indentation.

Indentation is rejected as inconsistent if a source file mixes tabs and spaces in a way that makes the meaning dependent on the worth of a tab in spaces; a TabError is raised in that case.

Cross-platform compatibility note: because of the nature of text editors on non-UNIX platforms, it is unwise to use a mixture of spaces and tabs for the indentation in a single source file. It should also be noted that different platforms may explicitly limit the maximum indentation level.

A formfeed character may be present at the start of the line; it will be ignored for the indentation calculations above. Formfeed characters occurring elsewhere in the leading whitespace have an undefined effect (for instance, they may reset the space count to zero).

The indentation levels of consecutive lines are used to generate INDENT and DEDENT tokens, using a stack, as follows.

Before the first line of the file is read, a single zero is pushed on the stack; this will never be popped off again. The numbers pushed on the stack will always be strictly increasing from bottom to top. At the beginning of each logical line, the line’s indentation level is compared to the top of the stack. If it is equal, nothing happens. If it is larger, it is pushed on the stack, and one INDENT token is generated. If it is smaller, it must be one of the numbers occurring on the stack; all numbers on the stack that are larger are popped off, and for each number popped off a DEDENT token is generated. At the end of the file, a DEDENT token is generated for each number remaining on the stack that is larger than zero.

Here is an example of a correctly (though confusingly) indented piece of Python code:

def perm(l):
        # Compute the list of all permutations of l
    if len(l) <= 1:
                  return [l]
    r = []
    for i in range(len(l)):
             s = l[:i] + l[i+1:]
             p = perm(s)
             for x in p:
              r.append(l[i:i+1] + x)
    return r

The following example shows various indentation errors:

 def perm(l):                       # error: first line indented
for i in range(len(l)):             # error: not indented
    s = l[:i] + l[i+1:]
        p = perm(l[:i] + l[i+1:])   # error: unexpected indent
        for x in p:
                r.append(l[i:i+1] + x)
            return r                # error: inconsistent dedent

(Actually, the first three errors are detected by the parser; only the last error is found by the lexical analyzer — the indentation of return r does not match a level popped off the stack.)

2.1.9. Whitespace between tokens

Except at the beginning of a logical line or in string literals, the whitespace characters space, tab and formfeed can be used interchangeably to separate tokens. Whitespace is needed between two tokens only if their concatenation could otherwise be interpreted as a different token (e.g., ab is one token, but a b is two tokens).

2.2. Other tokens

Besides NEWLINE, INDENT and DEDENT, the following categories of tokens exist: identifierskeywordsliteralsoperators, and delimiters. Whitespace characters (other than line terminators, discussed earlier) are not tokens, but serve to delimit tokens. Where ambiguity exists, a token comprises the longest possible string that forms a legal token, when read from left to right.

2.3. Identifiers and keywords

Identifiers (also referred to as names) are described by the following lexical definitions.

The syntax of identifiers in Python is based on the Unicode standard annex UAX-31, with elaboration and changes as defined below; see also PEP 3131 for further details.

Within the ASCII range (U+0001..U+007F), the valid characters for identifiers are the same as in Python 2.x: the uppercase and lowercase letters A through Z, the underscore _ and, except for the first character, the digits 0 through 9.

Python 3.0 introduces additional characters from outside the ASCII range (see PEP 3131). For these characters, the classification uses the version of the Unicode Character Database as included in the unicodedata module.

Identifiers are unlimited in length. Case is significant.

identifier   ::=  xid_start xid_continue* id_start  ::= id_continue  ::= id_start, plus characters in the categories Mn, Mc, Nd, Pc and others with the Other_ID_Continue property> xid_start  ::= id_start whose NFKC normalization is in "id_start xid_continue*"> xid_continue ::= id_continue whose NFKC normalization is in "id_continue*">

The Unicode category codes mentioned above stand for:

  • Lu – uppercase letters

  • Ll – lowercase letters

  • Lt – titlecase letters

  • Lm – modifier letters

  • Lo – other letters

  • Nl – letter numbers

  • Mn – nonspacing marks

  • Mc – spacing combining marks

  • Nd – decimal numbers

  • Pc – connector punctuations

  • Other_ID_Start – explicit list of characters in PropList.txt to support backwards compatibility

  • Other_ID_Continue – likewise

All identifiers are converted into the normal form NFKC while parsing; comparison of identifiers is based on NFKC.

A non-normative HTML file listing all valid identifier characters for Unicode 4.1 can be found at https://www.unicode.org/Public/13.0.0/ucd/DerivedCoreProperties.txt

2.3.1. Keywords

The following identifiers are used as reserved words, or keywords of the language, and cannot be used as ordinary identifiers. They must be spelled exactly as written here:

False      await      else       import     pass
None       break      except     in         raise
True       class      finally    is         return
and        continue   for        lambda     try
as         def        from       nonlocal   while
assert     del        global     not        with
async      elif       if         or         yield

2.3.2. Soft Keywords

New in version 3.10.

Some identifiers are only reserved under specific contexts. These are known as soft keywords. The identifiers matchcase and _ can syntactically act as keywords in contexts related to the pattern matching statement, but this distinction is done at the parser level, not when tokenizing.

As soft keywords, their use with pattern matching is possible while still preserving compatibility with existing code that uses matchcase and _ as identifier names.

2.3.3. Reserved classes of identifiers

Certain classes of identifiers (besides keywords) have special meanings. These classes are identified by the patterns of leading and trailing underscore characters:

_*

Not imported by from module import *.

_

In a case pattern within a match statement, _ is a soft keyword that denotes a wildcard.

Separately, the interactive interpreter makes the result of the last evaluation available in the variable _. (It is stored in the builtins module, alongside built-in functions like print.)

Elsewhere, _ is a regular identifier. It is often used to name “special” items, but it is not special to Python itself.

Note

The name _ is often used in conjunction with internationalization; refer to the documentation for the gettext module for more information on this convention.

It is also commonly used for unused variables.

__*__

System-defined names, informally known as “dunder” names. These names are defined by the interpreter and its implementation (including the standard library). Current system names are discussed in the Special method names section and elsewhere. More will likely be defined in future versions of Python. Any use of __*__ names, in any context, that does not follow explicitly documented use, is subject to breakage without warning.

__*

Class-private names. Names in this category, when used within the context of a class definition, are re-written to use a mangled form to help avoid name clashes between “private” attributes of base and derived classes. See section Identifiers (Names).

2.4. Literals

Literals are notations for constant values of some built-in types.

2.4.1. String and Bytes literals

String literals are described by the following lexical definitions:

stringliteral   ::=  [stringprefix](shortstring | longstring) stringprefix  ::= "r" | "u" | "R" | "U" | "f" | "F" | "fr" | "Fr" | "fR" | "FR" | "rf" | "rF" | "Rf" | "RF" shortstring  ::= "'" shortstringitem* "'" | '"' shortstringitem* '"' longstring  ::= "'''" longstringitem* "'''" | '"""' longstringitem* '"""' shortstringitem ::= shortstringchar | stringescapeseq longstringitem  ::= longstringchar | stringescapeseq shortstringchar ::= longstringchar  ::= stringescapeseq ::= "\"
bytesliteral   ::=  bytesprefix(shortbytes | longbytes) bytesprefix  ::= "b" | "B" | "br" | "Br" | "bR" | "BR" | "rb" | "rB" | "Rb" | "RB" shortbytes  ::= "'" shortbytesitem* "'" | '"' shortbytesitem* '"' longbytes  ::= "'''" longbytesitem* "'''" | '"""' longbytesitem* '"""' shortbytesitem ::= shortbyteschar | bytesescapeseq longbytesitem  ::= longbyteschar | bytesescapeseq shortbyteschar ::= longbyteschar  ::= bytesescapeseq ::= "\"

One syntactic restriction not indicated by these productions is that whitespace is not allowed between the stringprefix or bytesprefix and the rest of the literal. The source character set is defined by the encoding declaration; it is UTF-8 if no encoding declaration is given in the source file; see section Encoding declarations.

In plain English: Both types of literals can be enclosed in matching single quotes (') or double quotes ("). They can also be enclosed in matching groups of three single or double quotes (these are generally referred to as triple-quoted strings). The backslash (\) character is used to escape characters that otherwise have a special meaning, such as newline, backslash itself, or the quote character.

Bytes literals are always prefixed with 'b' or 'B'; they produce an instance of the bytes type instead of the str type. They may only contain ASCII characters; bytes with a numeric value of 128 or greater must be expressed with escapes.

Both string and bytes literals may optionally be prefixed with a letter 'r' or 'R'; such strings are called raw strings and treat backslashes as literal characters. As a result, in string literals, '\U' and '\u' escapes in raw strings are not treated specially. Given that Python 2.x’s raw unicode literals behave differently than Python 3.x’s the 'ur' syntax is not supported.

New in version 3.3: The 'rb' prefix of raw bytes literals has been added as a synonym of 'br'.

New in version 3.3: Support for the unicode legacy literal (u'value') was reintroduced to simplify the maintenance of dual Python 2.x and 3.x codebases. See PEP 414 for more information.

A string literal with 'f' or 'F' in its prefix is a formatted string literal; see Formatted string literals. The 'f' may be combined with 'r', but not with 'b' or 'u', therefore raw formatted strings are possible, but formatted bytes literals are not.

In triple-quoted literals, unescaped newlines and quotes are allowed (and are retained), except that three unescaped quotes in a row terminate the literal. (A “quote” is the character used to open the literal, i.e. either ' or ".)

Unless an 'r' or 'R' prefix is present, escape sequences in string and bytes literals are interpreted according to rules similar to those used by Standard C. The recognized escape sequences are:

Escape Sequence

Meaning

Notes

\newline

Backslash and newline ignored

 

\\

Backslash (\)

 

\'

Single quote (')

 

\"

Double quote (")

 

\a

ASCII Bell (BEL)

 

\b

ASCII Backspace (BS)

 

\f

ASCII Formfeed (FF)

 

\n

ASCII Linefeed (LF)

 

\r

ASCII Carriage Return (CR)

 

\t

ASCII Horizontal Tab (TAB)

 

\v

ASCII Vertical Tab (VT)

 

\ooo

Character with octal value ooo

(1,3)

\xhh

Character with hex value hh

(2,3)

Escape sequences only recognized in string literals are:

Escape Sequence

Meaning

Notes

\N{name}

Character named name in the Unicode database

(4)

\uxxxx

Character with 16-bit hex value xxxx

(5)

\Uxxxxxxxx

Character with 32-bit hex value xxxxxxxx

(6)

Notes:

  1. As in Standard C, up to three octal digits are accepted.

  2. Unlike in Standard C, exactly two hex digits are required.

  3. In a bytes literal, hexadecimal and octal escapes denote the byte with the given value. In a string literal, these escapes denote a Unicode character with the given value.

  4. Changed in version 3.3: Support for name aliases 1 has been added.

  5. Exactly four hex digits are required.

  6. Any Unicode character can be encoded this way. Exactly eight hex digits are required.

Unlike Standard C, all unrecognized escape sequences are left in the string unchanged, i.e., the backslash is left in the result. (This behavior is useful when debugging: if an escape sequence is mistyped, the resulting output is more easily recognized as broken.) It is also important to note that the escape sequences only recognized in string literals fall into the category of unrecognized escapes for bytes literals.

Changed in version 3.6: Unrecognized escape sequences produce a DeprecationWarning. In a future Python version they will be a SyntaxWarning and eventually a SyntaxError.

Even in a raw literal, quotes can be escaped with a backslash, but the backslash remains in the result; for example, r"\"" is a valid string literal consisting of two characters: a backslash and a double quote; r"\" is not a valid string literal (even a raw string cannot end in an odd number of backslashes). Specifically, a raw literal cannot end in a single backslash (since the backslash would escape the following quote character). Note also that a single backslash followed by a newline is interpreted as those two characters as part of the literal, not as a line continuation.

2.4.2. String literal concatenation

Multiple adjacent string or bytes literals (delimited by whitespace), possibly using different quoting conventions, are allowed, and their meaning is the same as their concatenation. Thus, "hello" 'world' is equivalent to "helloworld". This feature can be used to reduce the number of backslashes needed, to split long strings conveniently across long lines, or even to add comments to parts of strings, for example:

re.compile("[A-Za-z_]"       # letter or underscore
           "[A-Za-z0-9_]*"   # letter, digit or underscore
          )

Note that this feature is defined at the syntactical level, but implemented at compile time. The ‘+’ operator must be used to concatenate string expressions at run time. Also note that literal concatenation can use different quoting styles for each component (even mixing raw strings and triple quoted strings), and formatted string literals may be concatenated with plain string literals.

2.4.3. Formatted string literals

New in version 3.6.

formatted string literal or f-string is a string literal that is prefixed with 'f' or 'F'. These strings may contain replacement fields, which are expressions delimited by curly braces {}. While other string literals always have a constant value, formatted strings are really expressions evaluated at run time.

Escape sequences are decoded like in ordinary string literals (except when a literal is also marked as a raw string). After decoding, the grammar for the contents of the string is:

f_string          ::=  (literal_char | "{{" | "}}" | replacement_field)* replacement_field ::= "{" f_expression ["="] ["!" conversion] [":" format_spec] "}" f_expression  ::= (conditional_expression | "*" or_expr) ("," conditional_expression | "," "*" or_expr)* [","] | yield_expression conversion  ::= "s" | "r" | "a" format_spec  ::= (literal_char | NULL | replacement_field)* literal_char  ::=

The parts of the string outside curly braces are treated literally, except that any doubled curly braces '{{' or '}}' are replaced with the corresponding single curly brace. A single opening curly bracket '{' marks a replacement field, which starts with a Python expression. To display both the expression text and its value after evaluation, (useful in debugging), an equal sign '=' may be added after the expression. A conversion field, introduced by an exclamation point '!' may follow. A format specifier may also be appended, introduced by a colon ':'. A replacement field ends with a closing curly bracket '}'.

Expressions in formatted string literals are treated like regular Python expressions surrounded by parentheses, with a few exceptions. An empty expression is not allowed, and both lambda and assignment expressions := must be surrounded by explicit parentheses. Replacement expressions can contain line breaks (e.g. in triple-quoted strings), but they cannot contain comments. Each expression is evaluated in the context where the formatted string literal appears, in order from left to right.

Changed in version 3.7: Prior to Python 3.7, an await expression and comprehensions containing an async for clause were illegal in the expressions in formatted string literals due to a problem with the implementation.

When the equal sign '=' is provided, the output will have the expression text, the '=' and the evaluated value. Spaces after the opening brace '{', within the expression and after the '=' are all retained in the output. By default, the '=' causes the repr() of the expression to be provided, unless there is a format specified. When a format is specified it defaults to the str() of the expression unless a conversion '!r' is declared.

New in version 3.8: The equal sign '='.

If a conversion is specified, the result of evaluating the expression is converted before formatting. Conversion '!s' calls str() on the result, '!r' calls repr(), and '!a' calls ascii().

The result is then formatted using the format() protocol. The format specifier is passed to the __format__() method of the expression or conversion result. An empty string is passed when the format specifier is omitted. The formatted result is then included in the final value of the whole string.

Top-level format specifiers may include nested replacement fields. These nested fields may include their own conversion fields and format specifiers, but may not include more deeply-nested replacement fields. The format specifier mini-language is the same as that used by the str.format() method.

Formatted string literals may be concatenated, but replacement fields cannot be split across literals.

Some examples of formatted string literals:

>>>

>>> name = "Fred"
>>> f"He said his name is {name!r}."
"He said his name is 'Fred'."
>>> f"He said his name is {repr(name)}."  # repr() is equivalent to !r
"He said his name is 'Fred'."
>>> width = 10
>>> precision = 4
>>> value = decimal.Decimal("12.34567")
>>> f"result: {value:{width}.{precision}}"  # nested fields
'result:      12.35'
>>> today = datetime(year=2017, month=1, day=27)
>>> f"{today:%B %d, %Y}"  # using date format specifier
'January 27, 2017'
>>> f"{today=:%B %d, %Y}" # using date format specifier and debugging
'today=January 27, 2017'
>>> number = 1024
>>> f"{number:#0x}"  # using integer format specifier
'0x400'
>>> foo = "bar"
>>> f"{ foo = }" # preserves whitespace
" foo = 'bar'"
>>> line = "The mill's closed"
>>> f"{line = }"
'line = "The mill\'s closed"'
>>> f"{line = :20}"
"line = The mill's closed   "
>>> f"{line = !r:20}"
'line = "The mill\'s closed" '

A consequence of sharing the same syntax as regular string literals is that characters in the replacement fields must not conflict with the quoting used in the outer formatted string literal:

f"abc {a["x"]} def"    # error: outer string literal ended prematurely
f"abc {a['x']} def"    # workaround: use different quoting

Backslashes are not allowed in format expressions and will raise an error:

f"newline: {ord('\n')}"  # raises SyntaxError

To include a value in which a backslash escape is required, create a temporary variable.

>>>

>>> newline = ord('\n')
>>> f"newline: {newline}"
'newline: 10'

Formatted string literals cannot be used as docstrings, even if they do not include expressions.

>>>

>>> def foo():
...     f"Not a docstring"
...
>>> foo.__doc__ is None
True

See also PEP 498 for the proposal that added formatted string literals, and str.format(), which uses a related format string mechanism.

2.4.4. Numeric literals

There are three types of numeric literals: integers, floating point numbers, and imaginary numbers. There are no complex literals (complex numbers can be formed by adding a real number and an imaginary number).

Note that numeric literals do not include a sign; a phrase like -1 is actually an expression composed of the unary operator ‘-’ and the literal 1.

2.4.5. Integer literals

Integer literals are described by the following lexical definitions:

integer      ::=  decinteger | bininteger | octinteger | hexinteger decinteger  ::= nonzerodigit (["_"] digit)* | "0"+ (["_"] "0")* bininteger  ::= "0" ("b" | "B") (["_"] bindigit)+ octinteger  ::= "0" ("o" | "O") (["_"] octdigit)+ hexinteger  ::= "0" ("x" | "X") (["_"] hexdigit)+ nonzerodigit ::= "1"..."9" digit  ::= "0"..."9" bindigit  ::= "0" | "1" octdigit  ::= "0"..."7" hexdigit  ::= digit | "a"..."f" | "A"..."F"

There is no limit for the length of integer literals apart from what can be stored in available memory.

Underscores are ignored for determining the numeric value of the literal. They can be used to group digits for enhanced readability. One underscore can occur between digits, and after base specifiers like 0x.

Note that leading zeros in a non-zero decimal number are not allowed. This is for disambiguation with C-style octal literals, which Python used before version 3.0.

Some examples of integer literals:

7     2147483647                        0o177    0b100110111
3     79228162514264337593543950336     0o377    0xdeadbeef
      100_000_000_000                   0b_1110_0101

Changed in version 3.6: Underscores are now allowed for grouping purposes in literals.

2.4.6. Floating point literals

Floating point literals are described by the following lexical definitions:

floatnumber   ::=  pointfloat | exponentfloat pointfloat  ::= [digitpart] fraction | digitpart "." exponentfloat ::= (digitpart | pointfloat) exponent digitpart  ::= digit (["_"] digit)* fraction  ::= "." digitpart exponent  ::= ("e" | "E") ["+" | "-"] digitpart

Note that the integer and exponent parts are always interpreted using radix 10. For example, 077e010 is legal, and denotes the same number as 77e10. The allowed range of floating point literals is implementation-dependent. As in integer literals, underscores are supported for digit grouping.

Some examples of floating point literals:

3.14    10.    .001    1e100    3.14e-10    0e0    3.14_15_93

Changed in version 3.6: Underscores are now allowed for grouping purposes in literals.

2.4.7. Imaginary literals

Imaginary literals are described by the following lexical definitions:

imagnumber ::=  (floatnumber | digitpart) ("j" | "J")

An imaginary literal yields a complex number with a real part of 0.0. Complex numbers are represented as a pair of floating point numbers and have the same restrictions on their range. To create a complex number with a nonzero real part, add a floating point number to it, e.g., (3+4j). Some examples of imaginary literals:

3.14j   10.j    10j     .001j   1e100j   3.14e-10j   3.14_15_93j

2.5. Operators

The following tokens are operators:

+       -       *       **      /       //      %      @
<< >>      &       |       ^       ~       :=
< >       <= >=      ==      !=

2.6. Delimiters

The following tokens serve as delimiters in the grammar:

(       )       [       ]       {       }
,       :       .       ;       @       =       ->
+=      -=      *=      /=      //=     %=      @=
&=      |=      ^=      >>=     <<=     **=

The period can also occur in floating-point and imaginary literals. A sequence of three periods has a special meaning as an ellipsis literal. The second half of the list, the augmented assignment operators, serve lexically as delimiters, but also perform an operation.

The following printing ASCII characters have special meaning as part of other tokens or are otherwise significant to the lexical analyzer:

'       "       #       \

The following printing ASCII characters are not used in Python. Their occurrence outside string literals and comments is an unconditional error:

$       ?       `

Footnotes

1

https://www.unicode.org/Public/11.0.0/ucd/NameAliases.txt

2. Lexical analysis

A Python program is read by a parser. Input to the parser is a stream of tokens, generated by the lexical analyzer. This chapter describes how the lexical analyzer breaks a file into tokens.

Python reads program text as Unicode code points; the encoding of a source file can be given by an encoding declaration and defaults to UTF-8, see PEP 3120 for details. If the source file cannot be decoded, a SyntaxError is raised.

2.1. Line structure

A Python program is divided into a number of logical lines.

2.1.1. Logical lines

The end of a logical line is represented by the token NEWLINE. Statements cannot cross logical line boundaries except where NEWLINE is allowed by the syntax (e.g., between statements in compound statements). A logical line is constructed from one or more physical lines by following the explicit or implicit line joining rules.

2.1.2. Physical lines

A physical line is a sequence of characters terminated by an end-of-line sequence. In source files and strings, any of the standard platform line termination sequences can be used – the Unix form using ASCII LF (linefeed), the Windows form using the ASCII sequence CR LF (return followed by linefeed), or the old Macintosh form using the ASCII CR (return) character. All of these forms can be used equally, regardless of platform. The end of input also serves as an implicit terminator for the final physical line.

When embedding Python, source code strings should be passed to Python APIs using the standard C conventions for newline characters (the \n character, representing ASCII LF, is the line terminator).

2.1.3. Comments

A comment starts with a hash character (#) that is not part of a string literal, and ends at the end of the physical line. A comment signifies the end of the logical line unless the implicit line joining rules are invoked. Comments are ignored by the syntax.

2.1.4. Encoding declarations

If a comment in the first or second line of the Python script matches the regular expression coding[=:]\s*([-\w.]+), this comment is processed as an encoding declaration; the first group of this expression names the encoding of the source code file. The encoding declaration must appear on a line of its own. If it is the second line, the first line must also be a comment-only line. The recommended forms of an encoding expression are

# -*- coding:  -*-

which is recognized also by GNU Emacs, and

# vim:fileencoding=

which is recognized by Bram Moolenaar’s VIM.

If no encoding declaration is found, the default encoding is UTF-8. In addition, if the first bytes of the file are the UTF-8 byte-order mark (b'\xef\xbb\xbf'), the declared file encoding is UTF-8 (this is supported, among others, by Microsoft’s notepad).

If an encoding is declared, the encoding name must be recognized by Python (see Standard Encodings). The encoding is used for all lexical analysis, including string literals, comments and identifiers.

2.1.5. Explicit line joining

Two or more physical lines may be joined into logical lines using backslash characters (\), as follows: when a physical line ends in a backslash that is not part of a string literal or comment, it is joined with the following forming a single logical line, deleting the backslash and the following end-of-line character. For example:

if 1900 < year < 2100 and 1 <= month <= 12 \
   and 1 <= day <= 31 and 0 <= hour < 24 \
   and 0 <= minute < 60 and 0 <= second < 60:   # Looks like a valid date
        return 1

A line ending in a backslash cannot carry a comment. A backslash does not continue a comment. A backslash does not continue a token except for string literals (i.e., tokens other than string literals cannot be split across physical lines using a backslash). A backslash is illegal elsewhere on a line outside a string literal.

2.1.6. Implicit line joining

Expressions in parentheses, square brackets or curly braces can be split over more than one physical line without using backslashes. For example:

month_names = ['Januari', 'Februari', 'Maart',      # These are the
               'April',   'Mei',      'Juni',       # Dutch names
               'Juli',    'Augustus', 'September',  # for the months
               'Oktober', 'November', 'December']   # of the year

Implicitly continued lines can carry comments. The indentation of the continuation lines is not important. Blank continuation lines are allowed. There is no NEWLINE token between implicit continuation lines. Implicitly continued lines can also occur within triple-quoted strings (see below); in that case they cannot carry comments.

2.1.7. Blank lines

A logical line that contains only spaces, tabs, formfeeds and possibly a comment, is ignored (i.e., no NEWLINE token is generated). During interactive input of statements, handling of a blank line may differ depending on the implementation of the read-eval-print loop. In the standard interactive interpreter, an entirely blank logical line (i.e. one containing not even whitespace or a comment) terminates a multi-line statement.

2.1.8. Indentation

Leading whitespace (spaces and tabs) at the beginning of a logical line is used to compute the indentation level of the line, which in turn is used to determine the grouping of statements.

Tabs are replaced (from left to right) by one to eight spaces such that the total number of characters up to and including the replacement is a multiple of eight (this is intended to be the same rule as used by Unix). The total number of spaces preceding the first non-blank character then determines the line’s indentation. Indentation cannot be split over multiple physical lines using backslashes; the whitespace up to the first backslash determines the indentation.

Indentation is rejected as inconsistent if a source file mixes tabs and spaces in a way that makes the meaning dependent on the worth of a tab in spaces; a TabError is raised in that case.

Cross-platform compatibility note: because of the nature of text editors on non-UNIX platforms, it is unwise to use a mixture of spaces and tabs for the indentation in a single source file. It should also be noted that different platforms may explicitly limit the maximum indentation level.

A formfeed character may be present at the start of the line; it will be ignored for the indentation calculations above. Formfeed characters occurring elsewhere in the leading whitespace have an undefined effect (for instance, they may reset the space count to zero).

The indentation levels of consecutive lines are used to generate INDENT and DEDENT tokens, using a stack, as follows.

Before the first line of the file is read, a single zero is pushed on the stack; this will never be popped off again. The numbers pushed on the stack will always be strictly increasing from bottom to top. At the beginning of each logical line, the line’s indentation level is compared to the top of the stack. If it is equal, nothing happens. If it is larger, it is pushed on the stack, and one INDENT token is generated. If it is smaller, it must be one of the numbers occurring on the stack; all numbers on the stack that are larger are popped off, and for each number popped off a DEDENT token is generated. At the end of the file, a DEDENT token is generated for each number remaining on the stack that is larger than zero.

Here is an example of a correctly (though confusingly) indented piece of Python code:

def perm(l):
        # Compute the list of all permutations of l
    if len(l) <= 1:
                  return [l]
    r = []
    for i in range(len(l)):
             s = l[:i] + l[i+1:]
             p = perm(s)
             for x in p:
              r.append(l[i:i+1] + x)
    return r

The following example shows various indentation errors:

 def perm(l):                       # error: first line indented
for i in range(len(l)):             # error: not indented
    s = l[:i] + l[i+1:]
        p = perm(l[:i] + l[i+1:])   # error: unexpected indent
        for x in p:
                r.append(l[i:i+1] + x)
            return r                # error: inconsistent dedent

(Actually, the first three errors are detected by the parser; only the last error is found by the lexical analyzer — the indentation of return r does not match a level popped off the stack.)

2.1.9. Whitespace between tokens

Except at the beginning of a logical line or in string literals, the whitespace characters space, tab and formfeed can be used interchangeably to separate tokens. Whitespace is needed between two tokens only if their concatenation could otherwise be interpreted as a different token (e.g., ab is one token, but a b is two tokens).

2.2. Other tokens

Besides NEWLINE, INDENT and DEDENT, the following categories of tokens exist: identifierskeywordsliteralsoperators, and delimiters. Whitespace characters (other than line terminators, discussed earlier) are not tokens, but serve to delimit tokens. Where ambiguity exists, a token comprises the longest possible string that forms a legal token, when read from left to right.

2.3. Identifiers and keywords

Identifiers (also referred to as names) are described by the following lexical definitions.

The syntax of identifiers in Python is based on the Unicode standard annex UAX-31, with elaboration and changes as defined below; see also PEP 3131 for further details.

Within the ASCII range (U+0001..U+007F), the valid characters for identifiers are the same as in Python 2.x: the uppercase and lowercase letters A through Z, the underscore _ and, except for the first character, the digits 0 through 9.

Python 3.0 introduces additional characters from outside the ASCII range (see PEP 3131). For these characters, the classification uses the version of the Unicode Character Database as included in the unicodedata module.

Identifiers are unlimited in length. Case is significant.

identifier   ::=  xid_start xid_continue* id_start  ::= id_continue  ::= id_start, plus characters in the categories Mn, Mc, Nd, Pc and others with the Other_ID_Continue property> xid_start  ::= id_start whose NFKC normalization is in "id_start xid_continue*"> xid_continue ::= id_continue whose NFKC normalization is in "id_continue*">

The Unicode category codes mentioned above stand for:

  • Lu – uppercase letters

  • Ll – lowercase letters

  • Lt – titlecase letters

  • Lm – modifier letters

  • Lo – other letters

  • Nl – letter numbers

  • Mn – nonspacing marks

  • Mc – spacing combining marks

  • Nd – decimal numbers

  • Pc – connector punctuations

  • Other_ID_Start – explicit list of characters in PropList.txt to support backwards compatibility

  • Other_ID_Continue – likewise

All identifiers are converted into the normal form NFKC while parsing; comparison of identifiers is based on NFKC.

A non-normative HTML file listing all valid identifier characters for Unicode 4.1 can be found at https://www.unicode.org/Public/13.0.0/ucd/DerivedCoreProperties.txt

2.3.1. Keywords

The following identifiers are used as reserved words, or keywords of the language, and cannot be used as ordinary identifiers. They must be spelled exactly as written here:

False      await      else       import     pass
None       break      except     in         raise
True       class      finally    is         return
and        continue   for        lambda     try
as         def        from       nonlocal   while
assert     del        global     not        with
async      elif       if         or         yield

2.3.2. Soft Keywords

New in version 3.10.

Some identifiers are only reserved under specific contexts. These are known as soft keywords. The identifiers matchcase and _ can syntactically act as keywords in contexts related to the pattern matching statement, but this distinction is done at the parser level, not when tokenizing.

As soft keywords, their use with pattern matching is possible while still preserving compatibility with existing code that uses matchcase and _ as identifier names.

2.3.3. Reserved classes of identifiers

Certain classes of identifiers (besides keywords) have special meanings. These classes are identified by the patterns of leading and trailing underscore characters:

_*

Not imported by from module import *.

_

In a case pattern within a match statement, _ is a soft keyword that denotes a wildcard.

Separately, the interactive interpreter makes the result of the last evaluation available in the variable _. (It is stored in the builtins module, alongside built-in functions like print.)

Elsewhere, _ is a regular identifier. It is often used to name “special” items, but it is not special to Python itself.

Note

The name _ is often used in conjunction with internationalization; refer to the documentation for the gettext module for more information on this convention.

It is also commonly used for unused variables.

__*__

System-defined names, informally known as “dunder” names. These names are defined by the interpreter and its implementation (including the standard library). Current system names are discussed in the Special method names section and elsewhere. More will likely be defined in future versions of Python. Any use of __*__ names, in any context, that does not follow explicitly documented use, is subject to breakage without warning.

__*

Class-private names. Names in this category, when used within the context of a class definition, are re-written to use a mangled form to help avoid name clashes between “private” attributes of base and derived classes. See section Identifiers (Names).

2.4. Literals

Literals are notations for constant values of some built-in types.

2.4.1. String and Bytes literals

String literals are described by the following lexical definitions:

stringliteral   ::=  [stringprefix](shortstring | longstring) stringprefix  ::= "r" | "u" | "R" | "U" | "f" | "F" | "fr" | "Fr" | "fR" | "FR" | "rf" | "rF" | "Rf" | "RF" shortstring  ::= "'" shortstringitem* "'" | '"' shortstringitem* '"' longstring  ::= "'''" longstringitem* "'''" | '"""' longstringitem* '"""' shortstringitem ::= shortstringchar | stringescapeseq longstringitem  ::= longstringchar | stringescapeseq shortstringchar ::= longstringchar  ::= stringescapeseq ::= "\"
bytesliteral   ::=  bytesprefix(shortbytes | longbytes) bytesprefix  ::= "b" | "B" | "br" | "Br" | "bR" | "BR" | "rb" | "rB" | "Rb" | "RB" shortbytes  ::= "'" shortbytesitem* "'" | '"' shortbytesitem* '"' longbytes  ::= "'''" longbytesitem* "'''" | '"""' longbytesitem* '"""' shortbytesitem ::= shortbyteschar | bytesescapeseq longbytesitem  ::= longbyteschar | bytesescapeseq shortbyteschar ::= longbyteschar  ::= bytesescapeseq ::= "\"

One syntactic restriction not indicated by these productions is that whitespace is not allowed between the stringprefix or bytesprefix and the rest of the literal. The source character set is defined by the encoding declaration; it is UTF-8 if no encoding declaration is given in the source file; see section Encoding declarations.

In plain English: Both types of literals can be enclosed in matching single quotes (') or double quotes ("). They can also be enclosed in matching groups of three single or double quotes (these are generally referred to as triple-quoted strings). The backslash (\) character is used to escape characters that otherwise have a special meaning, such as newline, backslash itself, or the quote character.

Bytes literals are always prefixed with 'b' or 'B'; they produce an instance of the bytes type instead of the str type. They may only contain ASCII characters; bytes with a numeric value of 128 or greater must be expressed with escapes.

Both string and bytes literals may optionally be prefixed with a letter 'r' or 'R'; such strings are called raw strings and treat backslashes as literal characters. As a result, in string literals, '\U' and '\u' escapes in raw strings are not treated specially. Given that Python 2.x’s raw unicode literals behave differently than Python 3.x’s the 'ur' syntax is not supported.

New in version 3.3: The 'rb' prefix of raw bytes literals has been added as a synonym of 'br'.

New in version 3.3: Support for the unicode legacy literal (u'value') was reintroduced to simplify the maintenance of dual Python 2.x and 3.x codebases. See PEP 414 for more information.

A string literal with 'f' or 'F' in its prefix is a formatted string literal; see Formatted string literals. The 'f' may be combined with 'r', but not with 'b' or 'u', therefore raw formatted strings are possible, but formatted bytes literals are not.

In triple-quoted literals, unescaped newlines and quotes are allowed (and are retained), except that three unescaped quotes in a row terminate the literal. (A “quote” is the character used to open the literal, i.e. either ' or ".)

Unless an 'r' or 'R' prefix is present, escape sequences in string and bytes literals are interpreted according to rules similar to those used by Standard C. The recognized escape sequences are:

Escape Sequence

Meaning

Notes

\newline

Backslash and newline ignored

 

\\

Backslash (\)

 

\'

Single quote (')

 

\"

Double quote (")

 

\a

ASCII Bell (BEL)

 

\b

ASCII Backspace (BS)

 

\f

ASCII Formfeed (FF)

 

\n

ASCII Linefeed (LF)

 

\r

ASCII Carriage Return (CR)

 

\t

ASCII Horizontal Tab (TAB)

 

\v

ASCII Vertical Tab (VT)

 

\ooo

Character with octal value ooo

(1,3)

\xhh

Character with hex value hh

(2,3)

Escape sequences only recognized in string literals are:

Escape Sequence

Meaning

Notes

\N{name}

Character named name in the Unicode database

(4)

\uxxxx

Character with 16-bit hex value xxxx

(5)

\Uxxxxxxxx

Character with 32-bit hex value xxxxxxxx

(6)

Notes:

  1. As in Standard C, up to three octal digits are accepted.

  2. Unlike in Standard C, exactly two hex digits are required.

  3. In a bytes literal, hexadecimal and octal escapes denote the byte with the given value. In a string literal, these escapes denote a Unicode character with the given value.

  4. Changed in version 3.3: Support for name aliases 1 has been added.

  5. Exactly four hex digits are required.

  6. Any Unicode character can be encoded this way. Exactly eight hex digits are required.

Unlike Standard C, all unrecognized escape sequences are left in the string unchanged, i.e., the backslash is left in the result. (This behavior is useful when debugging: if an escape sequence is mistyped, the resulting output is more easily recognized as broken.) It is also important to note that the escape sequences only recognized in string literals fall into the category of unrecognized escapes for bytes literals.

Changed in version 3.6: Unrecognized escape sequences produce a DeprecationWarning. In a future Python version they will be a SyntaxWarning and eventually a SyntaxError.

Even in a raw literal, quotes can be escaped with a backslash, but the backslash remains in the result; for example, r"\"" is a valid string literal consisting of two characters: a backslash and a double quote; r"\" is not a valid string literal (even a raw string cannot end in an odd number of backslashes). Specifically, a raw literal cannot end in a single backslash (since the backslash would escape the following quote character). Note also that a single backslash followed by a newline is interpreted as those two characters as part of the literal, not as a line continuation.

2.4.2. String literal concatenation

Multiple adjacent string or bytes literals (delimited by whitespace), possibly using different quoting conventions, are allowed, and their meaning is the same as their concatenation. Thus, "hello" 'world' is equivalent to "helloworld". This feature can be used to reduce the number of backslashes needed, to split long strings conveniently across long lines, or even to add comments to parts of strings, for example:

re.compile("[A-Za-z_]"       # letter or underscore
           "[A-Za-z0-9_]*"   # letter, digit or underscore
          )

Note that this feature is defined at the syntactical level, but implemented at compile time. The ‘+’ operator must be used to concatenate string expressions at run time. Also note that literal concatenation can use different quoting styles for each component (even mixing raw strings and triple quoted strings), and formatted string literals may be concatenated with plain string literals.

2.4.3. Formatted string literals

New in version 3.6.

formatted string literal or f-string is a string literal that is prefixed with 'f' or 'F'. These strings may contain replacement fields, which are expressions delimited by curly braces {}. While other string literals always have a constant value, formatted strings are really expressions evaluated at run time.

Escape sequences are decoded like in ordinary string literals (except when a literal is also marked as a raw string). After decoding, the grammar for the contents of the string is:

f_string          ::=  (literal_char | "{{" | "}}" | replacement_field)* replacement_field ::= "{" f_expression ["="] ["!" conversion] [":" format_spec] "}" f_expression  ::= (conditional_expression | "*" or_expr) ("," conditional_expression | "," "*" or_expr)* [","] | yield_expression conversion  ::= "s" | "r" | "a" format_spec  ::= (literal_char | NULL | replacement_field)* literal_char  ::=

The parts of the string outside curly braces are treated literally, except that any doubled curly braces '{{' or '}}' are replaced with the corresponding single curly brace. A single opening curly bracket '{' marks a replacement field, which starts with a Python expression. To display both the expression text and its value after evaluation, (useful in debugging), an equal sign '=' may be added after the expression. A conversion field, introduced by an exclamation point '!' may follow. A format specifier may also be appended, introduced by a colon ':'. A replacement field ends with a closing curly bracket '}'.

Expressions in formatted string literals are treated like regular Python expressions surrounded by parentheses, with a few exceptions. An empty expression is not allowed, and both lambda and assignment expressions := must be surrounded by explicit parentheses. Replacement expressions can contain line breaks (e.g. in triple-quoted strings), but they cannot contain comments. Each expression is evaluated in the context where the formatted string literal appears, in order from left to right.

Changed in version 3.7: Prior to Python 3.7, an await expression and comprehensions containing an async for clause were illegal in the expressions in formatted string literals due to a problem with the implementation.

When the equal sign '=' is provided, the output will have the expression text, the '=' and the evaluated value. Spaces after the opening brace '{', within the expression and after the '=' are all retained in the output. By default, the '=' causes the repr() of the expression to be provided, unless there is a format specified. When a format is specified it defaults to the str() of the expression unless a conversion '!r' is declared.

New in version 3.8: The equal sign '='.

If a conversion is specified, the result of evaluating the expression is converted before formatting. Conversion '!s' calls str() on the result, '!r' calls repr(), and '!a' calls ascii().

The result is then formatted using the format() protocol. The format specifier is passed to the __format__() method of the expression or conversion result. An empty string is passed when the format specifier is omitted. The formatted result is then included in the final value of the whole string.

Top-level format specifiers may include nested replacement fields. These nested fields may include their own conversion fields and format specifiers, but may not include more deeply-nested replacement fields. The format specifier mini-language is the same as that used by the str.format() method.

Formatted string literals may be concatenated, but replacement fields cannot be split across literals.

Some examples of formatted string literals:

>>>

>>> name = "Fred"
>>> f"He said his name is {name!r}."
"He said his name is 'Fred'."
>>> f"He said his name is {repr(name)}."  # repr() is equivalent to !r
"He said his name is 'Fred'."
>>> width = 10
>>> precision = 4
>>> value = decimal.Decimal("12.34567")
>>> f"result: {value:{width}.{precision}}"  # nested fields
'result:      12.35'
>>> today = datetime(year=2017, month=1, day=27)
>>> f"{today:%B %d, %Y}"  # using date format specifier
'January 27, 2017'
>>> f"{today=:%B %d, %Y}" # using date format specifier and debugging
'today=January 27, 2017'
>>> number = 1024
>>> f"{number:#0x}"  # using integer format specifier
'0x400'
>>> foo = "bar"
>>> f"{ foo = }" # preserves whitespace
" foo = 'bar'"
>>> line = "The mill's closed"
>>> f"{line = }"
'line = "The mill\'s closed"'
>>> f"{line = :20}"
"line = The mill's closed   "
>>> f"{line = !r:20}"
'line = "The mill\'s closed" '

A consequence of sharing the same syntax as regular string literals is that characters in the replacement fields must not conflict with the quoting used in the outer formatted string literal:

f"abc {a["x"]} def"    # error: outer string literal ended prematurely
f"abc {a['x']} def"    # workaround: use different quoting

Backslashes are not allowed in format expressions and will raise an error:

f"newline: {ord('\n')}"  # raises SyntaxError

To include a value in which a backslash escape is required, create a temporary variable.

>>>

>>> newline = ord('\n')
>>> f"newline: {newline}"
'newline: 10'

Formatted string literals cannot be used as docstrings, even if they do not include expressions.

>>>

>>> def foo():
...     f"Not a docstring"
...
>>> foo.__doc__ is None
True

See also PEP 498 for the proposal that added formatted string literals, and str.format(), which uses a related format string mechanism.

2.4.4. Numeric literals

There are three types of numeric literals: integers, floating point numbers, and imaginary numbers. There are no complex literals (complex numbers can be formed by adding a real number and an imaginary number).

Note that numeric literals do not include a sign; a phrase like -1 is actually an expression composed of the unary operator ‘-’ and the literal 1.

2.4.5. Integer literals

Integer literals are described by the following lexical definitions:

integer      ::=  decinteger | bininteger | octinteger | hexinteger decinteger  ::= nonzerodigit (["_"] digit)* | "0"+ (["_"] "0")* bininteger  ::= "0" ("b" | "B") (["_"] bindigit)+ octinteger  ::= "0" ("o" | "O") (["_"] octdigit)+ hexinteger  ::= "0" ("x" | "X") (["_"] hexdigit)+ nonzerodigit ::= "1"..."9" digit  ::= "0"..."9" bindigit  ::= "0" | "1" octdigit  ::= "0"..."7" hexdigit  ::= digit | "a"..."f" | "A"..."F"

There is no limit for the length of integer literals apart from what can be stored in available memory.

Underscores are ignored for determining the numeric value of the literal. They can be used to group digits for enhanced readability. One underscore can occur between digits, and after base specifiers like 0x.

Note that leading zeros in a non-zero decimal number are not allowed. This is for disambiguation with C-style octal literals, which Python used before version 3.0.

Some examples of integer literals:

7     2147483647                        0o177    0b100110111
3     79228162514264337593543950336     0o377    0xdeadbeef
      100_000_000_000                   0b_1110_0101

Changed in version 3.6: Underscores are now allowed for grouping purposes in literals.

2.4.6. Floating point literals

Floating point literals are described by the following lexical definitions:

floatnumber   ::=  pointfloat | exponentfloat pointfloat  ::= [digitpart] fraction | digitpart "." exponentfloat ::= (digitpart | pointfloat) exponent digitpart  ::= digit (["_"] digit)* fraction  ::= "." digitpart exponent  ::= ("e" | "E") ["+" | "-"] digitpart

Note that the integer and exponent parts are always interpreted using radix 10. For example, 077e010 is legal, and denotes the same number as 77e10. The allowed range of floating point literals is implementation-dependent. As in integer literals, underscores are supported for digit grouping.

Some examples of floating point literals:

3.14    10.    .001    1e100    3.14e-10    0e0    3.14_15_93

Changed in version 3.6: Underscores are now allowed for grouping purposes in literals.

2.4.7. Imaginary literals

Imaginary literals are described by the following lexical definitions:

imagnumber ::=  (floatnumber | digitpart) ("j" | "J")

An imaginary literal yields a complex number with a real part of 0.0. Complex numbers are represented as a pair of floating point numbers and have the same restrictions on their range. To create a complex number with a nonzero real part, add a floating point number to it, e.g., (3+4j). Some examples of imaginary literals:

3.14j   10.j    10j     .001j   1e100j   3.14e-10j   3.14_15_93j

2.5. Operators

The following tokens are operators:

+       -       *       **      /       //      %      @
<< >>      &       |       ^       ~       :=
< >       <= >=      ==      !=

2.6. Delimiters

The following tokens serve as delimiters in the grammar:

(       )       [       ]       {       }
,       :       .       ;       @       =       ->
+=      -=      *=      /=      //=     %=      @=
&=      |=      ^=      >>=     <<=     **=

The period can also occur in floating-point and imaginary literals. A sequence of three periods has a special meaning as an ellipsis literal. The second half of the list, the augmented assignment operators, serve lexically as delimiters, but also perform an operation.

The following printing ASCII characters have special meaning as part of other tokens or are otherwise significant to the lexical analyzer:

'       "       #       \

The following printing ASCII characters are not used in Python. Their occurrence outside string literals and comments is an unconditional error:

$       ?       `

Footnotes

1

https://www.unicode.org/Public/11.0.0/ucd/NameAliases.txt

3. Data model

3.1. Objects, values and types

Objects are Python’s abstraction for data. All data in a Python program is represented by objects or by relations between objects. (In a sense, and in conformance to Von Neumann’s model of a “stored program computer”, code is also represented by objects.)

 

Every object has an identity, a type and a value. An object’s identity never changes once it has been created; you may think of it as the object’s address in memory. The ‘is’ operator compares the identity of two objects; the id() function returns an integer representing its identity.

CPython implementation detail: For CPython, id(x) is the memory address where x is stored.

An object’s type determines the operations that the object supports (e.g., “does it have a length?”) and also defines the possible values for objects of that type. The type() function returns an object’s type (which is an object itself). Like its identity, an object’s type is also unchangeable. 1

The value of some objects can change. Objects whose value can change are said to be mutable; objects whose value is unchangeable once they are created are called immutable. (The value of an immutable container object that contains a reference to a mutable object can change when the latter’s value is changed; however the container is still considered immutable, because the collection of objects it contains cannot be changed. So, immutability is not strictly the same as having an unchangeable value, it is more subtle.) An object’s mutability is determined by its type; for instance, numbers, strings and tuples are immutable, while dictionaries and lists are mutable.

Objects are never explicitly destroyed; however, when they become unreachable they may be garbage-collected. An implementation is allowed to postpone garbage collection or omit it altogether — it is a matter of implementation quality how garbage collection is implemented, as long as no objects are collected that are still reachable.

CPython implementation detail: CPython currently uses a reference-counting scheme with (optional) delayed detection of cyclically linked garbage, which collects most objects as soon as they become unreachable, but is not guaranteed to collect garbage containing circular references. See the documentation of the gc module for information on controlling the collection of cyclic garbage. Other implementations act differently and CPython may change. Do not depend on immediate finalization of objects when they become unreachable (so you should always close files explicitly).

Note that the use of the implementation’s tracing or debugging facilities may keep objects alive that would normally be collectable. Also note that catching an exception with a ‘tryexcept’ statement may keep objects alive.

Some objects contain references to “external” resources such as open files or windows. It is understood that these resources are freed when the object is garbage-collected, but since garbage collection is not guaranteed to happen, such objects also provide an explicit way to release the external resource, usually a close() method. Programs are strongly recommended to explicitly close such objects. The ‘tryfinally’ statement and the ‘with’ statement provide convenient ways to do this.

Some objects contain references to other objects; these are called containers. Examples of containers are tuples, lists and dictionaries. The references are part of a container’s value. In most cases, when we talk about the value of a container, we imply the values, not the identities of the contained objects; however, when we talk about the mutability of a container, only the identities of the immediately contained objects are implied. So, if an immutable container (like a tuple) contains a reference to a mutable object, its value changes if that mutable object is changed.

Types affect almost all aspects of object behavior. Even the importance of object identity is affected in some sense: for immutable types, operations that compute new values may actually return a reference to any existing object with the same type and value, while for mutable objects this is not allowed. E.g., after a = 1; b = 1a and b may or may not refer to the same object with the value one, depending on the implementation, but after c = []; d = []c and d are guaranteed to refer to two different, unique, newly created empty lists. (Note that c = d = [] assigns the same object to both c and d.)

3.2. The standard type hierarchy

Below is a list of the types that are built into Python. Extension modules (written in C, Java, or other languages, depending on the implementation) can define additional types. Future versions of Python may add types to the type hierarchy (e.g., rational numbers, efficiently stored arrays of integers, etc.), although such additions will often be provided via the standard library instead.

Some of the type descriptions below contain a paragraph listing ‘special attributes.’ These are attributes that provide access to the implementation and are not intended for general use. Their definition may change in the future.

None

This type has a single value. There is a single object with this value. This object is accessed through the built-in name None. It is used to signify the absence of a value in many situations, e.g., it is returned from functions that don’t explicitly return anything. Its truth value is false.

NotImplemented

This type has a single value. There is a single object with this value. This object is accessed through the built-in name NotImplemented. Numeric methods and rich comparison methods should return this value if they do not implement the operation for the operands provided. (The interpreter will then try the reflected operation, or some other fallback, depending on the operator.) It should not be evaluated in a boolean context.

See Implementing the arithmetic operations for more details.

Changed in version 3.9: Evaluating NotImplemented in a boolean context is deprecated. While it currently evaluates as true, it will emit a DeprecationWarning. It will raise a TypeError in a future version of Python.

Ellipsis

This type has a single value. There is a single object with this value. This object is accessed through the literal ... or the built-in name Ellipsis. Its truth value is true.

numbers.Number

These are created by numeric literals and returned as results by arithmetic operators and arithmetic built-in functions. Numeric objects are immutable; once created their value never changes. Python numbers are of course strongly related to mathematical numbers, but subject to the limitations of numerical representation in computers.

The string representations of the numeric classes, computed by __repr__() and __str__(), have the following properties:

  • They are valid numeric literals which, when passed to their class constructor, produce an object having the value of the original numeric.

  • The representation is in base 10, when possible.

  • Leading zeros, possibly excepting a single zero before a decimal point, are not shown.

  • Trailing zeros, possibly excepting a single zero after a decimal point, are not shown.

  • A sign is shown only when the number is negative.

Python distinguishes between integers, floating point numbers, and complex numbers:

numbers.Integral

These represent elements from the mathematical set of integers (positive and negative).

There are two types of integers:

Integers (int)

These represent numbers in an unlimited range, subject to available (virtual) memory only. For the purpose of shift and mask operations, a binary representation is assumed, and negative numbers are represented in a variant of 2’s complement which gives the illusion of an infinite string of sign bits extending to the left.

Booleans (bool)

These represent the truth values False and True. The two objects representing the values False and True are the only Boolean objects. The Boolean type is a subtype of the integer type, and Boolean values behave like the values 0 and 1, respectively, in almost all contexts, the exception being that when converted to a string, the strings "False" or "True" are returned, respectively.

The rules for integer representation are intended to give the most meaningful interpretation of shift and mask operations involving negative integers.

numbers.Real (float)

These represent machine-level double precision floating point numbers. You are at the mercy of the underlying machine architecture (and C or Java implementation) for the accepted range and handling of overflow. Python does not support single-precision floating point numbers; the savings in processor and memory usage that are usually the reason for using these are dwarfed by the overhead of using objects in Python, so there is no reason to complicate the language with two kinds of floating point numbers.

numbers.Complex (complex)

These represent complex numbers as a pair of machine-level double precision floating point numbers. The same caveats apply as for floating point numbers. The real and imaginary parts of a complex number z can be retrieved through the read-only attributes z.real and z.imag.

Sequences

These represent finite ordered sets indexed by non-negative numbers. The built-in function len() returns the number of items of a sequence. When the length of a sequence is n, the index set contains the numbers 0, 1, …, n-1. Item i of sequence a is selected by a[i].

Sequences also support slicing: a[i:j] selects all items with index k such that i <= k < j. When used as an expression, a slice is a sequence of the same type. This implies that the index set is renumbered so that it starts at 0.

Some sequences also support “extended slicing” with a third “step” parameter: a[i:j:k] selects all items of a with index x where x = i + n*kn >= 0 and i <= x < j.

Sequences are distinguished according to their mutability:

Immutable sequences

An object of an immutable sequence type cannot change once it is created. (If the object contains references to other objects, these other objects may be mutable and may be changed; however, the collection of objects directly referenced by an immutable object cannot change.)

The following types are immutable sequences:

Strings

A string is a sequence of values that represent Unicode code points. All the code points in the range U+0000 - U+10FFFF can be represented in a string. Python doesn’t have a char type; instead, every code point in the string is represented as a string object with length 1. The built-in function ord() converts a code point from its string form to an integer in the range 0 - 10FFFFchr() converts an integer in the range 0 - 10FFFF to the corresponding length 1 string object. str.encode() can be used to convert a str to bytes using the given text encoding, and bytes.decode() can be used to achieve the opposite.

Tuples

The items of a tuple are arbitrary Python objects. Tuples of two or more items are formed by comma-separated lists of expressions. A tuple of one item (a ‘singleton’) can be formed by affixing a comma to an expression (an expression by itself does not create a tuple, since parentheses must be usable for grouping of expressions). An empty tuple can be formed by an empty pair of parentheses.

Bytes

A bytes object is an immutable array. The items are 8-bit bytes, represented by integers in the range 0 <= x < 256. Bytes literals (like b'abc') and the built-in bytes() constructor can be used to create bytes objects. Also, bytes objects can be decoded to strings via the decode() method.

Mutable sequences

Mutable sequences can be changed after they are created. The subscription and slicing notations can be used as the target of assignment and del (delete) statements.

There are currently two intrinsic mutable sequence types:

Lists

The items of a list are arbitrary Python objects. Lists are formed by placing a comma-separated list of expressions in square brackets. (Note that there are no special cases needed to form lists of length 0 or 1.)

Byte Arrays

A bytearray object is a mutable array. They are created by the built-in bytearray() constructor. Aside from being mutable (and hence unhashable), byte arrays otherwise provide the same interface and functionality as immutable bytes objects.

The extension module array provides an additional example of a mutable sequence type, as does the collections module.

Set types

These represent unordered, finite sets of unique, immutable objects. As such, they cannot be indexed by any subscript. However, they can be iterated over, and the built-in function len() returns the number of items in a set. Common uses for sets are fast membership testing, removing duplicates from a sequence, and computing mathematical operations such as intersection, union, difference, and symmetric difference.

For set elements, the same immutability rules apply as for dictionary keys. Note that numeric types obey the normal rules for numeric comparison: if two numbers compare equal (e.g., 1 and 1.0), only one of them can be contained in a set.

There are currently two intrinsic set types:

Sets

These represent a mutable set. They are created by the built-in set() constructor and can be modified afterwards by several methods, such as add().

Frozen sets

These represent an immutable set. They are created by the built-in frozenset() constructor. As a frozenset is immutable and hashable, it can be used again as an element of another set, or as a dictionary key.

Mappings

These represent finite sets of objects indexed by arbitrary index sets. The subscript notation a[k] selects the item indexed by k from the mapping a; this can be used in expressions and as the target of assignments or del statements. The built-in function len() returns the number of items in a mapping.

There is currently a single intrinsic mapping type:

Dictionaries

These represent finite sets of objects indexed by nearly arbitrary values. The only types of values not acceptable as keys are values containing lists or dictionaries or other mutable types that are compared by value rather than by object identity, the reason being that the efficient implementation of dictionaries requires a key’s hash value to remain constant. Numeric types used for keys obey the normal rules for numeric comparison: if two numbers compare equal (e.g., 1 and 1.0) then they can be used interchangeably to index the same dictionary entry.

Dictionaries preserve insertion order, meaning that keys will be produced in the same order they were added sequentially over the dictionary. Replacing an existing key does not change the order, however removing a key and re-inserting it will add it to the end instead of keeping its old place.

Dictionaries are mutable; they can be created by the {...} notation (see section Dictionary displays).

The extension modules dbm.ndbm and dbm.gnu provide additional examples of mapping types, as does the collections module.

Changed in version 3.7: Dictionaries did not preserve insertion order in versions of Python before 3.6. In CPython 3.6, insertion order was preserved, but it was considered an implementation detail at that time rather than a language guarantee.

Callable types

These are the types to which the function call operation (see section Calls) can be applied:

User-defined functions

A user-defined function object is created by a function definition (see section Function definitions). It should be called with an argument list containing the same number of items as the function’s formal parameter list.

Special attributes:

Attribute

Meaning

 

__doc__

The function’s documentation string, or None if unavailable; not inherited by subclasses.

Writable

__name__

The function’s name.

Writable

__qualname__

The function’s qualified name.

New in version 3.3.

Writable

__module__

The name of the module the function was defined in, or None if unavailable.

Writable

__defaults__

A tuple containing default argument values for those arguments that have defaults, or None if no arguments have a default value.

Writable

__code__

The code object representing the compiled function body.

Writable

__globals__

A reference to the dictionary that holds the function’s global variables — the global namespace of the module in which the function was defined.

Read-only

__dict__

The namespace supporting arbitrary function attributes.

Writable

__closure__

None or a tuple of cells that contain bindings for the function’s free variables. See below for information on the cell_contents attribute.

Read-only

__annotations__

A dict containing annotations of parameters. The keys of the dict are the parameter names, and 'return' for the return annotation, if provided. For more information on working with this attribute, see Annotations Best Practices.

Writable

__kwdefaults__

A dict containing defaults for keyword-only parameters.

Writable

Most of the attributes labelled “Writable” check the type of the assigned value.

Function objects also support getting and setting arbitrary attributes, which can be used, for example, to attach metadata to functions. Regular attribute dot-notation is used to get and set such attributes. Note that the current implementation only supports function attributes on user-defined functions. Function attributes on built-in functions may be supported in the future.

A cell object has the attribute cell_contents. This can be used to get the value of the cell, as well as set the value.

Additional information about a function’s definition can be retrieved from its code object; see the description of internal types below. The cell type can be accessed in the types module.

Instance methods

An instance method object combines a class, a class instance and any callable object (normally a user-defined function).

Special read-only attributes: __self__ is the class instance object, __func__ is the function object; __doc__ is the method’s documentation (same as __func__.__doc__); __name__ is the method name (same as __func__.__name__); __module__ is the name of the module the method was defined in, or None if unavailable.

Methods also support accessing (but not setting) the arbitrary function attributes on the underlying function object.

User-defined method objects may be created when getting an attribute of a class (perhaps via an instance of that class), if that attribute is a user-defined function object or a class method object.

When an instance method object is created by retrieving a user-defined function object from a class via one of its instances, its __self__ attribute is the instance, and the method object is said to be bound. The new method’s __func__ attribute is the original function object.

When an instance method object is created by retrieving a class method object from a class or instance, its __self__ attribute is the class itself, and its __func__ attribute is the function object underlying the class method.

When an instance method object is called, the underlying function (__func__) is called, inserting the class instance (__self__) in front of the argument list. For instance, when C is a class which contains a definition for a function f(), and x is an instance of C, calling x.f(1) is equivalent to calling C.f(x, 1).

When an instance method object is derived from a class method object, the “class instance” stored in __self__ will actually be the class itself, so that calling either x.f(1) or C.f(1) is equivalent to calling f(C,1) where f is the underlying function.

Note that the transformation from function object to instance method object happens each time the attribute is retrieved from the instance. In some cases, a fruitful optimization is to assign the attribute to a local variable and call that local variable. Also notice that this transformation only happens for user-defined functions; other callable objects (and all non-callable objects) are retrieved without transformation. It is also important to note that user-defined functions which are attributes of a class instance are not converted to bound methods; this only happens when the function is an attribute of the class.

Generator functions

A function or method which uses the yield statement (see section The yield statement) is called a generator function. Such a function, when called, always returns an iterator object which can be used to execute the body of the function: calling the iterator’s iterator.__next__() method will cause the function to execute until it provides a value using the yield statement. When the function executes a return statement or falls off the end, a StopIteration exception is raised and the iterator will have reached the end of the set of values to be returned.

Coroutine functions

A function or method which is defined using async def is called a coroutine function. Such a function, when called, returns a coroutine object. It may contain await expressions, as well as async with and async for statements. See also the Coroutine Objects section.

Asynchronous generator functions

A function or method which is defined using async def and which uses the yield statement is called a asynchronous generator function. Such a function, when called, returns an asynchronous iterator object which can be used in an async for statement to execute the body of the function.

Calling the asynchronous iterator’s aiterator.__anext__ method will return an awaitable which when awaited will execute until it provides a value using the yield expression. When the function executes an empty return statement or falls off the end, a StopAsyncIteration exception is raised and the asynchronous iterator will have reached the end of the set of values to be yielded.

Built-in functions

A built-in function object is a wrapper around a C function. Examples of built-in functions are len() and math.sin() (math is a standard built-in module). The number and type of the arguments are determined by the C function. Special read-only attributes: __doc__ is the function’s documentation string, or None if unavailable; __name__ is the function’s name; __self__ is set to None (but see the next item); __module__ is the name of the module the function was defined in or None if unavailable.

Built-in methods

This is really a different disguise of a built-in function, this time containing an object passed to the C function as an implicit extra argument. An example of a built-in method is alist.append(), assuming alist is a list object. In this case, the special read-only attribute __self__ is set to the object denoted by alist.

Classes

Classes are callable. These objects normally act as factories for new instances of themselves, but variations are possible for class types that override __new__(). The arguments of the call are passed to __new__() and, in the typical case, to __init__() to initialize the new instance.

Class Instances

Instances of arbitrary classes can be made callable by defining a __call__() method in their class.

Modules

Modules are a basic organizational unit of Python code, and are created by the import system as invoked either by the import statement, or by calling functions such as importlib.import_module() and built-in __import__(). A module object has a namespace implemented by a dictionary object (this is the dictionary referenced by the __globals__ attribute of functions defined in the module). Attribute references are translated to lookups in this dictionary, e.g., m.x is equivalent to m.__dict__["x"]. A module object does not contain the code object used to initialize the module (since it isn’t needed once the initialization is done).

Attribute assignment updates the module’s namespace dictionary, e.g., m.x = 1 is equivalent to m.__dict__["x"] = 1.

Predefined (writable) attributes:

__name__

The module’s name.

__doc__

The module’s documentation string, or None if unavailable.

__file__

The pathname of the file from which the module was loaded, if it was loaded from a file. The __file__ attribute may be missing for certain types of modules, such as C modules that are statically linked into the interpreter. For extension modules loaded dynamically from a shared library, it’s the pathname of the shared library file.

__annotations__

A dictionary containing variable annotations collected during module body execution. For best practices on working with __annotations__, please see Annotations Best Practices.

Special read-only attribute: __dict__ is the module’s namespace as a dictionary object.

CPython implementation detail: Because of the way CPython clears module dictionaries, the module dictionary will be cleared when the module falls out of scope even if the dictionary still has live references. To avoid this, copy the dictionary or keep the module around while using its dictionary directly.

Custom classes

Custom class types are typically created by class definitions (see section Class definitions). A class has a namespace implemented by a dictionary object. Class attribute references are translated to lookups in this dictionary, e.g., C.x is translated to C.__dict__["x"] (although there are a number of hooks which allow for other means of locating attributes). When the attribute name is not found there, the attribute search continues in the base classes. This search of the base classes uses the C3 method resolution order which behaves correctly even in the presence of ‘diamond’ inheritance structures where there are multiple inheritance paths leading back to a common ancestor. Additional details on the C3 MRO used by Python can be found in the documentation accompanying the 2.3 release at https://www.python.org/download/releases/2.3/mro/.

When a class attribute reference (for class C, say) would yield a class method object, it is transformed into an instance method object whose __self__ attribute is C. When it would yield a static method object, it is transformed into the object wrapped by the static method object. See section Implementing Descriptors for another way in which attributes retrieved from a class may differ from those actually contained in its __dict__.

Class attribute assignments update the class’s dictionary, never the dictionary of a base class.

A class object can be called (see above) to yield a class instance (see below).

Special attributes:

__name__

The class name.

__module__

The name of the module in which the class was defined.

__dict__

The dictionary containing the class’s namespace.

__bases__

A tuple containing the base classes, in the order of their occurrence in the base class list.

__doc__

The class’s documentation string, or None if undefined.

__annotations__

A dictionary containing variable annotations collected during class body execution. For best practices on working with __annotations__, please see Annotations Best Practices.

Class instances

A class instance is created by calling a class object (see above). A class instance has a namespace implemented as a dictionary which is the first place in which attribute references are searched. When an attribute is not found there, and the instance’s class has an attribute by that name, the search continues with the class attributes. If a class attribute is found that is a user-defined function object, it is transformed into an instance method object whose __self__ attribute is the instance. Static method and class method objects are also transformed; see above under “Classes”. See section Implementing Descriptors for another way in which attributes of a class retrieved via its instances may differ from the objects actually stored in the class’s __dict__. If no class attribute is found, and the object’s class has a __getattr__() method, that is called to satisfy the lookup.

Attribute assignments and deletions update the instance’s dictionary, never a class’s dictionary. If the class has a __setattr__() or __delattr__() method, this is called instead of updating the instance dictionary directly.

Class instances can pretend to be numbers, sequences, or mappings if they have methods with certain special names. See section Special method names.

Special attributes: __dict__ is the attribute dictionary; __class__ is the instance’s class.

I/O objects (also known as file objects)

file object represents an open file. Various shortcuts are available to create file objects: the open() built-in function, and also os.popen()os.fdopen(), and the makefile() method of socket objects (and perhaps by other functions or methods provided by extension modules).

The objects sys.stdinsys.stdout and sys.stderr are initialized to file objects corresponding to the interpreter’s standard input, output and error streams; they are all open in text mode and therefore follow the interface defined by the io.TextIOBase abstract class.

Internal types

A few types used internally by the interpreter are exposed to the user. Their definitions may change with future versions of the interpreter, but they are mentioned here for completeness.

Code objects

Code objects represent byte-compiled executable Python code, or bytecode. The difference between a code object and a function object is that the function object contains an explicit reference to the function’s globals (the module in which it was defined), while a code object contains no context; also the default argument values are stored in the function object, not in the code object (because they represent values calculated at run-time). Unlike function objects, code objects are immutable and contain no references (directly or indirectly) to mutable objects.

Special read-only attributes: co_name gives the function name; co_argcount is the total number of positional arguments (including positional-only arguments and arguments with default values); co_posonlyargcount is the number of positional-only arguments (including arguments with default values); co_kwonlyargcount is the number of keyword-only arguments (including arguments with default values); co_nlocals is the number of local variables used by the function (including arguments); co_varnames is a tuple containing the names of the local variables (starting with the argument names); co_cellvars is a tuple containing the names of local variables that are referenced by nested functions; co_freevars is a tuple containing the names of free variables; co_code is a string representing the sequence of bytecode instructions; co_consts is a tuple containing the literals used by the bytecode; co_names is a tuple containing the names used by the bytecode; co_filename is the filename from which the code was compiled; co_firstlineno is the first line number of the function; co_lnotab is a string encoding the mapping from bytecode offsets to line numbers (for details see the source code of the interpreter); co_stacksize is the required stack size; co_flags is an integer encoding a number of flags for the interpreter.

The following flag bits are defined for co_flags: bit 0x04 is set if the function uses the *arguments syntax to accept an arbitrary number of positional arguments; bit 0x08 is set if the function uses the **keywords syntax to accept arbitrary keyword arguments; bit 0x20 is set if the function is a generator.

Future feature declarations (from __future__ import division) also use bits in co_flags to indicate whether a code object was compiled with a particular feature enabled: bit 0x2000 is set if the function was compiled with future division enabled; bits 0x10 and 0x1000 were used in earlier versions of Python.

Other bits in co_flags are reserved for internal use.

If a code object represents a function, the first item in co_consts is the documentation string of the function, or None if undefined.

Frame objects

Frame objects represent execution frames. They may occur in traceback objects (see below), and are also passed to registered trace functions.

Special read-only attributes: f_back is to the previous stack frame (towards the caller), or None if this is the bottom stack frame; f_code is the code object being executed in this frame; f_locals is the dictionary used to look up local variables; f_globals is used for global variables; f_builtins is used for built-in (intrinsic) names; f_lasti gives the precise instruction (this is an index into the bytecode string of the code object).

Accessing f_code raises an auditing event object.__getattr__ with arguments obj and "f_code".

Special writable attributes: f_trace, if not None, is a function called for various events during code execution (this is used by the debugger). Normally an event is triggered for each new source line – this can be disabled by setting f_trace_lines to False.

Implementations may allow per-opcode events to be requested by setting f_trace_opcodes to True. Note that this may lead to undefined interpreter behaviour if exceptions raised by the trace function escape to the function being traced.

f_lineno is the current line number of the frame — writing to this from within a trace function jumps to the given line (only for the bottom-most frame). A debugger can implement a Jump command (aka Set Next Statement) by writing to f_lineno.

Frame objects support one method:

frame.clear()

This method clears all references to local variables held by the frame. Also, if the frame belonged to a generator, the generator is finalized. This helps break reference cycles involving frame objects (for example when catching an exception and storing its traceback for later use).

RuntimeError is raised if the frame is currently executing.

New in version 3.4.

Traceback objects

Traceback objects represent a stack trace of an exception. A traceback object is implicitly created when an exception occurs, and may also be explicitly created by calling types.TracebackType.

For implicitly created tracebacks, when the search for an exception handler unwinds the execution stack, at each unwound level a traceback object is inserted in front of the current traceback. When an exception handler is entered, the stack trace is made available to the program. (See section The try statement.) It is accessible as the third item of the tuple returned by sys.exc_info(), and as the __traceback__ attribute of the caught exception.

When the program contains no suitable handler, the stack trace is written (nicely formatted) to the standard error stream; if the interpreter is interactive, it is also made available to the user as sys.last_traceback.

For explicitly created tracebacks, it is up to the creator of the traceback to determine how the tb_next attributes should be linked to form a full stack trace.

Special read-only attributes: tb_frame points to the execution frame of the current level; tb_lineno gives the line number where the exception occurred; tb_lasti indicates the precise instruction. The line number and last instruction in the traceback may differ from the line number of its frame object if the exception occurred in a try statement with no matching except clause or with a finally clause.

Accessing tb_frame raises an auditing event object.__getattr__ with arguments obj and "tb_frame".

Special writable attribute: tb_next is the next level in the stack trace (towards the frame where the exception occurred), or None if there is no next level.

Changed in version 3.7: Traceback objects can now be explicitly instantiated from Python code, and the tb_next attribute of existing instances can be updated.

Slice objects

Slice objects are used to represent slices for __getitem__() methods. They are also created by the built-in slice() function.

Special read-only attributes: start is the lower bound; stop is the upper bound; step is the step value; each is None if omitted. These attributes can have any type.

Slice objects support one method:

slice.indices(selflength)

This method takes a single integer argument length and computes information about the slice that the slice object would describe if applied to a sequence of length items. It returns a tuple of three integers; respectively these are the start and stop indices and the step or stride length of the slice. Missing or out-of-bounds indices are handled in a manner consistent with regular slices.

Static method objects

Static method objects provide a way of defeating the transformation of function objects to method objects described above. A static method object is a wrapper around any other object, usually a user-defined method object. When a static method object is retrieved from a class or a class instance, the object actually returned is the wrapped object, which is not subject to any further transformation. Static method objects are also callable. Static method objects are created by the built-in staticmethod() constructor.

Class method objects

A class method object, like a static method object, is a wrapper around another object that alters the way in which that object is retrieved from classes and class instances. The behaviour of class method objects upon such retrieval is described above, under “User-defined methods”. Class method objects are created by the built-in classmethod() constructor.

3.3. Special method names

A class can implement certain operations that are invoked by special syntax (such as arithmetic operations or subscripting and slicing) by defining methods with special names. This is Python’s approach to operator overloading, allowing classes to define their own behavior with respect to language operators. For instance, if a class defines a method named __getitem__(), and x is an instance of this class, then x[i] is roughly equivalent to type(x).__getitem__(x, i). Except where mentioned, attempts to execute an operation raise an exception when no appropriate method is defined (typically AttributeError or TypeError).

Setting a special method to None indicates that the corresponding operation is not available. For example, if a class sets __iter__() to None, the class is not iterable, so calling iter() on its instances will raise a TypeError (without falling back to __getitem__()). 2

When implementing a class that emulates any built-in type, it is important that the emulation only be implemented to the degree that it makes sense for the object being modelled. For example, some sequences may work well with retrieval of individual elements, but extracting a slice may not make sense. (One example of this is the NodeList interface in the W3C’s Document Object Model.)

3.3.1. Basic customization

object.__new__(cls[])

Called to create a new instance of class cls__new__() is a static method (special-cased so you need not declare it as such) that takes the class of which an instance was requested as its first argument. The remaining arguments are those passed to the object constructor expression (the call to the class). The return value of __new__() should be the new object instance (usually an instance of cls).

Typical implementations create a new instance of the class by invoking the superclass’s __new__() method using super().__new__(cls[, ...]) with appropriate arguments and then modifying the newly-created instance as necessary before returning it.

If __new__() is invoked during object construction and it returns an instance of cls, then the new instance’s __init__() method will be invoked like __init__(self[, ...]), where self is the new instance and the remaining arguments are the same as were passed to the object constructor.

If __new__() does not return an instance of cls, then the new instance’s __init__() method will not be invoked.

__new__() is intended mainly to allow subclasses of immutable types (like int, str, or tuple) to customize instance creation. It is also commonly overridden in custom metaclasses in order to customize class creation.

object.__init__(self[])

Called after the instance has been created (by __new__()), but before it is returned to the caller. The arguments are those passed to the class constructor expression. If a base class has an __init__() method, the derived class’s __init__() method, if any, must explicitly call it to ensure proper initialization of the base class part of the instance; for example: super().__init__([args...]).

Because __new__() and __init__() work together in constructing objects (__new__() to create it, and __init__() to customize it), no non-None value may be returned by __init__(); doing so will cause a TypeError to be raised at runtime.

object.__del__(self)

Called when the instance is about to be destroyed. This is also called a finalizer or (improperly) a destructor. If a base class has a __del__() method, the derived class’s __del__() method, if any, must explicitly call it to ensure proper deletion of the base class part of the instance.

It is possible (though not recommended!) for the __del__() method to postpone destruction of the instance by creating a new reference to it. This is called object resurrection. It is implementation-dependent whether __del__() is called a second time when a resurrected object is about to be destroyed; the current CPython implementation only calls it once.

It is not guaranteed that __del__() methods are called for objects that still exist when the interpreter exits.

Note

del x doesn’t directly call x.__del__() — the former decrements the reference count for x by one, and the latter is only called when x’s reference count reaches zero.

CPython implementation detail: It is possible for a reference cycle to prevent the reference count of an object from going to zero. In this case, the cycle will be later detected and deleted by the cyclic garbage collector. A common cause of reference cycles is when an exception has been caught in a local variable. The frame’s locals then reference the exception, which references its own traceback, which references the locals of all frames caught in the traceback.

See also

Documentation for the gc module.

Warning

Due to the precarious circumstances under which __del__() methods are invoked, exceptions that occur during their execution are ignored, and a warning is printed to sys.stderr instead. In particular:

  • __del__() can be invoked when arbitrary code is being executed, including from any arbitrary thread. If __del__() needs to take a lock or invoke any other blocking resource, it may deadlock as the resource may already be taken by the code that gets interrupted to execute __del__().

  • __del__() can be executed during interpreter shutdown. As a consequence, the global variables it needs to access (including other modules) may already have been deleted or set to None. Python guarantees that globals whose name begins with a single underscore are deleted from their module before other globals are deleted; if no other references to such globals exist, this may help in assuring that imported modules are still available at the time when the __del__() method is called.

 

object.__repr__(self)

Called by the repr() built-in function to compute the “official” string representation of an object. If at all possible, this should look like a valid Python expression that could be used to recreate an object with the same value (given an appropriate environment). If this is not possible, a string of the form <...some useful description...> should be returned. The return value must be a string object. If a class defines __repr__() but not __str__(), then __repr__() is also used when an “informal” string representation of instances of that class is required.

This is typically used for debugging, so it is important that the representation is information-rich and unambiguous.

 

object.__str__(self)

Called by str(object) and the built-in functions format() and print() to compute the “informal” or nicely printable string representation of an object. The return value must be a string object.

This method differs from object.__repr__() in that there is no expectation that __str__() return a valid Python expression: a more convenient or concise representation can be used.

The default implementation defined by the built-in type object calls object.__repr__().

object.__bytes__(self)

Called by bytes to compute a byte-string representation of an object. This should return a bytes object.

 

object.__format__(selfformat_spec)

Called by the format() built-in function, and by extension, evaluation of formatted string literals and the str.format() method, to produce a “formatted” string representation of an object. The format_spec argument is a string that contains a description of the formatting options desired. The interpretation of the format_spec argument is up to the type implementing __format__(), however most classes will either delegate formatting to one of the built-in types, or use a similar formatting option syntax.

See Format Specification Mini-Language for a description of the standard formatting syntax.

The return value must be a string object.

Changed in version 3.4: The __format__ method of object itself raises a TypeError if passed any non-empty string.

Changed in version 3.7: object.__format__(x, '') is now equivalent to str(x) rather than format(str(x), '').

 

object.__lt__(selfother)
object.__le__(selfother)
object.__eq__(selfother)
object.__ne__(selfother)
object.__gt__(selfother)
object.__ge__(selfother)

These are the so-called “rich comparison” methods. The correspondence between operator symbols and method names is as follows: x<y calls x.__lt__(y)x<=y calls x.__le__(y)x==y calls x.__eq__(y)x!=y calls x.__ne__(y)x>y calls x.__gt__(y), and x>=y calls x.__ge__(y).

A rich comparison method may return the singleton NotImplemented if it does not implement the operation for a given pair of arguments. By convention, False and True are returned for a successful comparison. However, these methods can return any value, so if the comparison operator is used in a Boolean context (e.g., in the condition of an if statement), Python will call bool() on the value to determine if the result is true or false.

By default, object implements __eq__() by using is, returning NotImplemented in the case of a false comparison: True if x is y else NotImplemented. For __ne__(), by default it delegates to __eq__() and inverts the result unless it is NotImplemented. There are no other implied relationships among the comparison operators or default implementations; for example, the truth of (x<y or x==y) does not imply x<=y. To automatically generate ordering operations from a single root operation, see functools.total_ordering().

See the paragraph on __hash__() for some important notes on creating hashable objects which support custom comparison operations and are usable as dictionary keys.

There are no swapped-argument versions of these methods (to be used when the left argument does not support the operation but the right argument does); rather, __lt__() and __gt__() are each other’s reflection, __le__() and __ge__() are each other’s reflection, and __eq__() and __ne__() are their own reflection. If the operands are of different types, and right operand’s type is a direct or indirect subclass of the left operand’s type, the reflected method of the right operand has priority, otherwise the left operand’s method has priority. Virtual subclassing is not considered.

object.__hash__(self)

Called by built-in function hash() and for operations on members of hashed collections including setfrozenset, and dict. The __hash__() method should return an integer. The only required property is that objects which compare equal have the same hash value; it is advised to mix together the hash values of the components of the object that also play a part in comparison of objects by packing them into a tuple and hashing the tuple. Example:

def __hash__(self):
    return hash((self.name, self.nick, self.color))

Note

hash() truncates the value returned from an object’s custom __hash__() method to the size of a Py_ssize_t. This is typically 8 bytes on 64-bit builds and 4 bytes on 32-bit builds. If an object’s __hash__() must interoperate on builds of different bit sizes, be sure to check the width on all supported builds. An easy way to do this is with python -c "import sys; print(sys.hash_info.width)".

If a class does not define an __eq__() method it should not define a __hash__() operation either; if it defines __eq__() but not __hash__(), its instances will not be usable as items in hashable collections. If a class defines mutable objects and implements an __eq__() method, it should not implement __hash__(), since the implementation of hashable collections requires that a key’s hash value is immutable (if the object’s hash value changes, it will be in the wrong hash bucket).

User-defined classes have __eq__() and __hash__() methods by default; with them, all objects compare unequal (except with themselves) and x.__hash__() returns an appropriate value such that x == y implies both that x is y and hash(x) == hash(y).

A class that overrides __eq__() and does not define __hash__() will have its __hash__() implicitly set to None. When the __hash__() method of a class is None, instances of the class will raise an appropriate TypeError when a program attempts to retrieve their hash value, and will also be correctly identified as unhashable when checking isinstance(obj, collections.abc.Hashable).

If a class that overrides __eq__() needs to retain the implementation of __hash__() from a parent class, the interpreter must be told this explicitly by setting __hash__ = .__hash__.

If a class that does not override __eq__() wishes to suppress hash support, it should include __hash__ = None in the class definition. A class which defines its own __hash__() that explicitly raises a TypeError would be incorrectly identified as hashable by an isinstance(obj, collections.abc.Hashable) call.

Note

By default, the __hash__() values of str and bytes objects are “salted” with an unpredictable random value. Although they remain constant within an individual Python process, they are not predictable between repeated invocations of Python.

This is intended to provide protection against a denial-of-service caused by carefully-chosen inputs that exploit the worst case performance of a dict insertion, O(n2) complexity. See http://www.ocert.org/advisories/ocert-2011-003.html for details.

Changing hash values affects the iteration order of sets. Python has never made guarantees about this ordering (and it typically varies between 32-bit and 64-bit builds).

See also PYTHONHASHSEED.

Changed in version 3.3: Hash randomization is enabled by default.

object.__bool__(self)

Called to implement truth value testing and the built-in operation bool(); should return False or True. When this method is not defined, __len__() is called, if it is defined, and the object is considered true if its result is nonzero. If a class defines neither __len__() nor __bool__(), all its instances are considered true.

3.3.2. Customizing attribute access

The following methods can be defined to customize the meaning of attribute access (use of, assignment to, or deletion of x.name) for class instances.

object.__getattr__(selfname)

Called when the default attribute access fails with an AttributeError (either __getattribute__() raises an AttributeError because name is not an instance attribute or an attribute in the class tree for self; or __get__() of a name property raises AttributeError). This method should either return the (computed) attribute value or raise an AttributeError exception.

Note that if the attribute is found through the normal mechanism, __getattr__() is not called. (This is an intentional asymmetry between __getattr__() and __setattr__().) This is done both for efficiency reasons and because otherwise __getattr__() would have no way to access other attributes of the instance. Note that at least for instance variables, you can fake total control by not inserting any values in the instance attribute dictionary (but instead inserting them in another object). See the __getattribute__() method below for a way to actually get total control over attribute access.

object.__getattribute__(selfname)

Called unconditionally to implement attribute accesses for instances of the class. If the class also defines __getattr__(), the latter will not be called unless __getattribute__() either calls it explicitly or raises an AttributeError. This method should return the (computed) attribute value or raise an AttributeError exception. In order to avoid infinite recursion in this method, its implementation should always call the base class method with the same name to access any attributes it needs, for example, object.__getattribute__(self, name).

Note

This method may still be bypassed when looking up special methods as the result of implicit invocation via language syntax or built-in functions. See Special method lookup.

For certain sensitive attribute accesses, raises an auditing event object.__getattr__ with arguments obj and name.

object.__setattr__(selfnamevalue)

Called when an attribute assignment is attempted. This is called instead of the normal mechanism (i.e. store the value in the instance dictionary). name is the attribute name, value is the value to be assigned to it.

If __setattr__() wants to assign to an instance attribute, it should call the base class method with the same name, for example, object.__setattr__(self, name, value).

For certain sensitive attribute assignments, raises an auditing event object.__setattr__ with arguments objnamevalue.

object.__delattr__(selfname)

Like __setattr__() but for attribute deletion instead of assignment. This should only be implemented if del obj.name is meaningful for the object.

For certain sensitive attribute deletions, raises an auditing event object.__delattr__ with arguments obj and name.

object.__dir__(self)

Called when dir() is called on the object. A sequence must be returned. dir() converts the returned sequence to a list and sorts it.

3.3.2.1. Customizing module attribute access

Special names __getattr__ and __dir__ can be also used to customize access to module attributes. The __getattr__ function at the module level should accept one argument which is the name of an attribute and return the computed value or raise an AttributeError. If an attribute is not found on a module object through the normal lookup, i.e. object.__getattribute__(), then __getattr__ is searched in the module __dict__ before raising an AttributeError. If found, it is called with the attribute name and the result is returned.

The __dir__ function should accept no arguments, and return a sequence of strings that represents the names accessible on module. If present, this function overrides the standard dir() search on a module.

For a more fine grained customization of the module behavior (setting attributes, properties, etc.), one can set the __class__ attribute of a module object to a subclass of types.ModuleType. For example:

import sys
from types import ModuleType

class VerboseModule(ModuleType):
    def __repr__(self):
        return f'Verbose {self.__name__}'

    def __setattr__(self, attr, value):
        print(f'Setting {attr}...')
        super().__setattr__(attr, value)

sys.modules[__name__].__class__ = VerboseModule

Note

Defining module __getattr__ and setting module __class__ only affect lookups made using the attribute access syntax – directly accessing the module globals (whether by code within the module, or via a reference to the module’s globals dictionary) is unaffected.

Changed in version 3.5: __class__ module attribute is now writable.

New in version 3.7: __getattr__ and __dir__ module attributes.

See also

PEP 562 – Module __getattr__ and __dir__

Describes the __getattr__ and __dir__ functions on modules.

3.3.2.2. Implementing Descriptors

The following methods only apply when an instance of the class containing the method (a so-called descriptor class) appears in an owner class (the descriptor must be in either the owner’s class dictionary or in the class dictionary for one of its parents). In the examples below, “the attribute” refers to the attribute whose name is the key of the property in the owner class’ __dict__.

object.__get__(selfinstanceowner=None)

Called to get the attribute of the owner class (class attribute access) or of an instance of that class (instance attribute access). The optional owner argument is the owner class, while instance is the instance that the attribute was accessed through, or None when the attribute is accessed through the owner.

This method should return the computed attribute value or raise an AttributeError exception.

PEP 252 specifies that __get__() is callable with one or two arguments. Python’s own built-in descriptors support this specification; however, it is likely that some third-party tools have descriptors that require both arguments. Python’s own __getattribute__() implementation always passes in both arguments whether they are required or not.

object.__set__(selfinstancevalue)

Called to set the attribute on an instance instance of the owner class to a new value, value.

Note, adding __set__() or __delete__() changes the kind of descriptor to a “data descriptor”. See Invoking Descriptors for more details.

object.__delete__(selfinstance)

Called to delete the attribute on an instance instance of the owner class.

The attribute __objclass__ is interpreted by the inspect module as specifying the class where this object was defined (setting this appropriately can assist in runtime introspection of dynamic class attributes). For callables, it may indicate that an instance of the given type (or a subclass) is expected or required as the first positional argument (for example, CPython sets this attribute for unbound methods that are implemented in C).

3.3.2.3. Invoking Descriptors

In general, a descriptor is an object attribute with “binding behavior”, one whose attribute access has been overridden by methods in the descriptor protocol: __get__()__set__(), and __delete__(). If any of those methods are defined for an object, it is said to be a descriptor.

The default behavior for attribute access is to get, set, or delete the attribute from an object’s dictionary. For instance, a.x has a lookup chain starting with a.__dict__['x'], then type(a).__dict__['x'], and continuing through the base classes of type(a) excluding metaclasses.

However, if the looked-up value is an object defining one of the descriptor methods, then Python may override the default behavior and invoke the descriptor method instead. Where this occurs in the precedence chain depends on which descriptor methods were defined and how they were called.

The starting point for descriptor invocation is a binding, a.x. How the arguments are assembled depends on a:

Direct Call

The simplest and least common call is when user code directly invokes a descriptor method: x.__get__(a).

Instance Binding

If binding to an object instance, a.x is transformed into the call: type(a).__dict__['x'].__get__(a, type(a)).

Class Binding

If binding to a class, A.x is transformed into the call: A.__dict__['x'].__get__(None, A).

Super Binding

If a is an instance of super, then the binding super(B, obj).m() searches obj.__class__.__mro__ for the base class A immediately following B and then invokes the descriptor with the call: A.__dict__['m'].__get__(obj, obj.__class__).

For instance bindings, the precedence of descriptor invocation depends on which descriptor methods are defined. A descriptor can define any combination of __get__()__set__() and __delete__(). If it does not define __get__(), then accessing the attribute will return the descriptor object itself unless there is a value in the object’s instance dictionary. If the descriptor defines __set__() and/or __delete__(), it is a data descriptor; if it defines neither, it is a non-data descriptor. Normally, data descriptors define both __get__() and __set__(), while non-data descriptors have just the __get__() method. Data descriptors with __get__() and __set__() (and/or __delete__()) defined always override a redefinition in an instance dictionary. In contrast, non-data descriptors can be overridden by instances.

Python methods (including those decorated with @staticmethod and @classmethod) are implemented as non-data descriptors. Accordingly, instances can redefine and override methods. This allows individual instances to acquire behaviors that differ from other instances of the same class.

The property() function is implemented as a data descriptor. Accordingly, instances cannot override the behavior of a property.

3.3.2.4. __slots__

__slots__ allow us to explicitly declare data members (like properties) and deny the creation of __dict__ and __weakref__ (unless explicitly declared in __slots__ or available in a parent.)

The space saved over using __dict__ can be significant. Attribute lookup speed can be significantly improved as well.

object.__slots__

This class variable can be assigned a string, iterable, or sequence of strings with variable names used by instances. __slots__ reserves space for the declared variables and prevents the automatic creation of __dict__ and __weakref__ for each instance.

3.3.2.4.1. Notes on using __slots__
  • When inheriting from a class without __slots__, the __dict__ and __weakref__ attribute of the instances will always be accessible.

  • Without a __dict__ variable, instances cannot be assigned new variables not listed in the __slots__ definition. Attempts to assign to an unlisted variable name raises AttributeError. If dynamic assignment of new variables is desired, then add '__dict__' to the sequence of strings in the __slots__ declaration.

  • Without a __weakref__ variable for each instance, classes defining __slots__ do not support weak references to its instances. If weak reference support is needed, then add '__weakref__' to the sequence of strings in the __slots__ declaration.

  • __slots__ are implemented at the class level by creating descriptors for each variable name. As a result, class attributes cannot be used to set default values for instance variables defined by __slots__; otherwise, the class attribute would overwrite the descriptor assignment.

  • The action of a __slots__ declaration is not limited to the class where it is defined. __slots__ declared in parents are available in child classes. However, child subclasses will get a __dict__ and __weakref__ unless they also define __slots__ (which should only contain names of any additional slots).

  • If a class defines a slot also defined in a base class, the instance variable defined by the base class slot is inaccessible (except by retrieving its descriptor directly from the base class). This renders the meaning of the program undefined. In the future, a check may be added to prevent this.

  • Nonempty __slots__ does not work for classes derived from “variable-length” built-in types such as intbytes and tuple.

  • Any non-string iterable may be assigned to __slots__.

  • If a dictionary is used to assign __slots__, the dictionary keys will be used as the slot names. The values of the dictionary can be used to provide per-attribute docstrings that will be recognised by inspect.getdoc() and displayed in the output of help().

  • __class__ assignment works only if both classes have the same __slots__.

  • Multiple inheritance with multiple slotted parent classes can be used, but only one parent is allowed to have attributes created by slots (the other bases must have empty slot layouts) – violations raise TypeError.

  • If an iterator is used for __slots__ then a descriptor is created for each of the iterator’s values. However, the __slots__ attribute will be an empty iterator.

3.3.3. Customizing class creation

Whenever a class inherits from another class, __init_subclass__() is called on the parent class. This way, it is possible to write classes which change the behavior of subclasses. This is closely related to class decorators, but where class decorators only affect the specific class they’re applied to, __init_subclass__ solely applies to future subclasses of the class defining the method.

classmethod object.__init_subclass__(cls)

This method is called whenever the containing class is subclassed. cls is then the new subclass. If defined as a normal instance method, this method is implicitly converted to a class method.

Keyword arguments which are given to a new class are passed to the parent’s class __init_subclass__. For compatibility with other classes using __init_subclass__, one should take out the needed keyword arguments and pass the others over to the base class, as in:

class Philosopher:
    def __init_subclass__(cls, /, default_name, **kwargs):
        super().__init_subclass__(**kwargs)
        cls.default_name = default_name

class AustralianPhilosopher(Philosopher, default_name="Bruce"):
    pass

The default implementation object.__init_subclass__ does nothing, but raises an error if it is called with any arguments.

Note

The metaclass hint metaclass is consumed by the rest of the type machinery, and is never passed to __init_subclass__ implementations. The actual metaclass (rather than the explicit hint) can be accessed as type(cls).

New in version 3.6.

When a class is created, type.__new__() scans the class variables and makes callbacks to those with a __set_name__() hook.

object.__set_name__(selfownername)

Automatically called at the time the owning class owner is created. The object has been assigned to name in that class:

class A:
    x = C()  # Automatically calls: x.__set_name__(A, 'x')

If the class variable is assigned after the class is created, __set_name__() will not be called automatically. If needed, __set_name__() can be called directly:

class A:
   pass

c = C()
A.x = c                  # The hook is not called
c.__set_name__(A, 'x')   # Manually invoke the hook

See Creating the class object for more details.

New in version 3.6.

3.3.3.1. Metaclasses

By default, classes are constructed using type(). The class body is executed in a new namespace and the class name is bound locally to the result of type(name, bases, namespace).

The class creation process can be customized by passing the metaclass keyword argument in the class definition line, or by inheriting from an existing class that included such an argument. In the following example, both MyClass and MySubclass are instances of Meta:

class Meta(type):
    pass

class MyClass(metaclass=Meta):
    pass

class MySubclass(MyClass):
    pass

Any other keyword arguments that are specified in the class definition are passed through to all metaclass operations described below.

When a class definition is executed, the following steps occur:

  • MRO entries are resolved;

  • the appropriate metaclass is determined;

  • the class namespace is prepared;

  • the class body is executed;

  • the class object is created.

3.3.3.2. Resolving MRO entries

If a base that appears in class definition is not an instance of type, then an __mro_entries__ method is searched on it. If found, it is called with the original bases tuple. This method must return a tuple of classes that will be used instead of this base. The tuple may be empty, in such case the original base is ignored.

See also

PEP 560 – Core support for typing module and generic types

3.3.3.3. Determining the appropriate metaclass

The appropriate metaclass for a class definition is determined as follows:

  • if no bases and no explicit metaclass are given, then type() is used;

  • if an explicit metaclass is given and it is not an instance of type(), then it is used directly as the metaclass;

  • if an instance of type() is given as the explicit metaclass, or bases are defined, then the most derived metaclass is used.

The most derived metaclass is selected from the explicitly specified metaclass (if any) and the metaclasses (i.e. type(cls)) of all specified base classes. The most derived metaclass is one which is a subtype of all of these candidate metaclasses. If none of the candidate metaclasses meets that criterion, then the class definition will fail with TypeError.

3.3.3.4. Preparing the class namespace

Once the appropriate metaclass has been identified, then the class namespace is prepared. If the metaclass has a __prepare__ attribute, it is called as namespace = metaclass.__prepare__(name, bases, **kwds) (where the additional keyword arguments, if any, come from the class definition). The __prepare__ method should be implemented as a classmethod. The namespace returned by __prepare__ is passed in to __new__, but when the final class object is created the namespace is copied into a new dict.

If the metaclass has no __prepare__ attribute, then the class namespace is initialised as an empty ordered mapping.

See also

PEP 3115 – Metaclasses in Python 3000

Introduced the __prepare__ namespace hook

3.3.3.5. Executing the class body

The class body is executed (approximately) as exec(body, globals(), namespace). The key difference from a normal call to exec() is that lexical scoping allows the class body (including any methods) to reference names from the current and outer scopes when the class definition occurs inside a function.

However, even when the class definition occurs inside the function, methods defined inside the class still cannot see names defined at the class scope. Class variables must be accessed through the first parameter of instance or class methods, or through the implicit lexically scoped __class__ reference described in the next section.

3.3.3.6. Creating the class object

Once the class namespace has been populated by executing the class body, the class object is created by calling metaclass(name, bases, namespace, **kwds) (the additional keywords passed here are the same as those passed to __prepare__).

This class object is the one that will be referenced by the zero-argument form of super()__class__ is an implicit closure reference created by the compiler if any methods in a class body refer to either __class__ or super. This allows the zero argument form of super() to correctly identify the class being defined based on lexical scoping, while the class or instance that was used to make the current call is identified based on the first argument passed to the method.

CPython implementation detail: In CPython 3.6 and later, the __class__ cell is passed to the metaclass as a __classcell__ entry in the class namespace. If present, this must be propagated up to the type.__new__ call in order for the class to be initialised correctly. Failing to do so will result in a RuntimeError in Python 3.8.

When using the default metaclass type, or any metaclass that ultimately calls type.__new__, the following additional customization steps are invoked after creating the class object:

  1. The type.__new__ method collects all of the attributes in the class namespace that define a __set_name__() method;

  2. Those __set_name__ methods are called with the class being defined and the assigned name of that particular attribute;

  3. The __init_subclass__() hook is called on the immediate parent of the new class in its method resolution order.

After the class object is created, it is passed to the class decorators included in the class definition (if any) and the resulting object is bound in the local namespace as the defined class.

When a new class is created by type.__new__, the object provided as the namespace parameter is copied to a new ordered mapping and the original object is discarded. The new copy is wrapped in a read-only proxy, which becomes the __dict__ attribute of the class object.

See also

PEP 3135 – New super

Describes the implicit __class__ closure reference

3.3.3.7. Uses for metaclasses

The potential uses for metaclasses are boundless. Some ideas that have been explored include enum, logging, interface checking, automatic delegation, automatic property creation, proxies, frameworks, and automatic resource locking/synchronization.

3.3.4. Customizing instance and subclass checks

The following methods are used to override the default behavior of the isinstance() and issubclass() built-in functions.

In particular, the metaclass abc.ABCMeta implements these methods in order to allow the addition of Abstract Base Classes (ABCs) as “virtual base classes” to any class or type (including built-in types), including other ABCs.

class.__instancecheck__(selfinstance)

Return true if instance should be considered a (direct or indirect) instance of class. If defined, called to implement isinstance(instance, class).

class.__subclasscheck__(selfsubclass)

Return true if subclass should be considered a (direct or indirect) subclass of class. If defined, called to implement issubclass(subclass, class).

Note that these methods are looked up on the type (metaclass) of a class. They cannot be defined as class methods in the actual class. This is consistent with the lookup of special methods that are called on instances, only in this case the instance is itself a class.

See also

PEP 3119 – Introducing Abstract Base Classes

Includes the specification for customizing isinstance() and issubclass() behavior through __instancecheck__() and __subclasscheck__(), with motivation for this functionality in the context of adding Abstract Base Classes (see the abc module) to the language.

3.3.5. Emulating generic types

When using type annotations, it is often useful to parameterize a generic type using Python’s square-brackets notation. For example, the annotation list[int] might be used to signify a list in which all the elements are of type int.

See also

PEP 484 – Type Hints

Introducing Python’s framework for type annotations

Generic Alias Types

Documentation for objects representing parameterized generic classes

Genericsuser-defined generics and typing.Generic

Documentation on how to implement generic classes that can be parameterized at runtime and understood by static type-checkers.

A class can generally only be parameterized if it defines the special class method __class_getitem__().

classmethod object.__class_getitem__(clskey)

Return an object representing the specialization of a generic class by type arguments found in key.

When defined on a class, __class_getitem__() is automatically a class method. As such, there is no need for it to be decorated with @classmethod when it is defined.

3.3.5.1. The purpose of __class_getitem__

The purpose of __class_getitem__() is to allow runtime parameterization of standard-library generic classes in order to more easily apply type hints to these classes.

To implement custom generic classes that can be parameterized at runtime and understood by static type-checkers, users should either inherit from a standard library class that already implements __class_getitem__(), or inherit from typing.Generic, which has its own implementation of __class_getitem__().

Custom implementations of __class_getitem__() on classes defined outside of the standard library may not be understood by third-party type-checkers such as mypy. Using __class_getitem__() on any class for purposes other than type hinting is discouraged.

3.3.5.2. __class_getitem__ versus __getitem__

Usually, the subscription of an object using square brackets will call the __getitem__() instance method defined on the object’s class. However, if the object being subscribed is itself a class, the class method __class_getitem__() may be called instead. __class_getitem__() should return a GenericAlias object if it is properly defined.

Presented with the expression obj[x], the Python interpreter follows something like the following process to decide whether __getitem__() or __class_getitem__() should be called:

from inspect import isclass

def subscribe(obj, x):
    """Return the result of the expression `obj[x]`"""

    class_of_obj = type(obj)

    # If the class of obj defines __getitem__,
    # call class_of_obj.__getitem__(obj, x)
    if hasattr(class_of_obj, '__getitem__'):
        return class_of_obj.__getitem__(obj, x)

    # Else, if obj is a class and defines __class_getitem__,
    # call obj.__class_getitem__(x)
    elif isclass(obj) and hasattr(obj, '__class_getitem__'):
        return obj.__class_getitem__(x)

    # Else, raise an exception
    else:
        raise TypeError(
            f"'{class_of_obj.__name__}' object is not subscriptable"
        )

In Python, all classes are themselves instances of other classes. The class of a class is known as that class’s metaclass, and most classes have the type class as their metaclass. type does not define __getitem__(), meaning that expressions such as list[int]dict[str, float] and tuple[str, bytes] all result in __class_getitem__() being called:

>>>

>>> # list has class "type" as its metaclass, like most classes:
>>> type(list)
<class 'type'>
>>> type(dict) == type(list) == type(tuple) == type(str) == type(bytes)
True
>>> # "list[int]" calls "list.__class_getitem__(int)"
>>> list[int]
list[int]
>>> # list.__class_getitem__ returns a GenericAlias object:
>>> type(list[int])
<class 'types.GenericAlias'>

However, if a class has a custom metaclass that defines __getitem__(), subscribing the class may result in different behaviour. An example of this can be found in the enum module:

>>>

>>> from enum import Enum
>>> class Menu(Enum):
...     """A breakfast menu"""
...     SPAM = 'spam'
...     BACON = 'bacon'
...
>>> # Enum classes have a custom metaclass:
>>> type(Menu)
<class 'enum.EnumMeta'>
>>> # EnumMeta defines __getitem__,
>>> # so __class_getitem__ is not called,
>>> # and the result is not a GenericAlias object:
>>> Menu['SPAM']
<Menu.SPAM: 'spam'>
>>> type(Menu['SPAM'])
<enum 'Menu'>

See also

PEP 560 – Core Support for typing module and generic types

Introducing __class_getitem__(), and outlining when a subscription results in __class_getitem__() being called instead of __getitem__()

3.3.6. Emulating callable objects

object.__call__(self[args…])

Called when the instance is “called” as a function; if this method is defined, x(arg1, arg2, ...) roughly translates to type(x).__call__(x, arg1, ...).

3.3.7. Emulating container types

The following methods can be defined to implement container objects. Containers usually are sequences (such as lists or tuples) or mappings (like dictionaries), but can represent other containers as well. The first set of methods is used either to emulate a sequence or to emulate a mapping; the difference is that for a sequence, the allowable keys should be the integers k for which 0 <= k < N where N is the length of the sequence, or slice objects, which define a range of items. It is also recommended that mappings provide the methods keys()values()items()get()clear()setdefault()pop()popitem()copy(), and update() behaving similar to those for Python’s standard dictionary objects. The collections.abc module provides a MutableMapping abstract base class to help create those methods from a base set of __getitem__()__setitem__()__delitem__(), and keys(). Mutable sequences should provide methods append()count()index()extend()insert()pop()remove()reverse() and sort(), like Python standard list objects. Finally, sequence types should implement addition (meaning concatenation) and multiplication (meaning repetition) by defining the methods __add__()__radd__()__iadd__()__mul__()__rmul__() and __imul__() described below; they should not define other numerical operators. It is recommended that both mappings and sequences implement the __contains__() method to allow efficient use of the in operator; for mappings, in should search the mapping’s keys; for sequences, it should search through the values. It is further recommended that both mappings and sequences implement the __iter__() method to allow efficient iteration through the container; for mappings, __iter__() should iterate through the object’s keys; for sequences, it should iterate through the values.

object.__len__(self)

Called to implement the built-in function len(). Should return the length of the object, an integer >= 0. Also, an object that doesn’t define a __bool__() method and whose __len__() method returns zero is considered to be false in a Boolean context.

CPython implementation detail: In CPython, the length is required to be at most sys.maxsize. If the length is larger than sys.maxsize some features (such as len()) may raise OverflowError. To prevent raising OverflowError by truth value testing, an object must define a __bool__() method.

object.__length_hint__(self)

Called to implement operator.length_hint(). Should return an estimated length for the object (which may be greater or less than the actual length). The length must be an integer >= 0. The return value may also be NotImplemented, which is treated the same as if the __length_hint__ method didn’t exist at all. This method is purely an optimization and is never required for correctness.

New in version 3.4.

Note

Slicing is done exclusively with the following three methods. A call like

a[1:2] = b

is translated to

a[slice(1, 2, None)] = b

and so forth. Missing slice items are always filled in with None.

object.__getitem__(selfkey)

Called to implement evaluation of self[key]. For sequence types, the accepted keys should be integers and slice objects. Note that the special interpretation of negative indexes (if the class wishes to emulate a sequence type) is up to the __getitem__() method. If key is of an inappropriate type, TypeError may be raised; if of a value outside the set of indexes for the sequence (after any special interpretation of negative values), IndexError should be raised. For mapping types, if key is missing (not in the container), KeyError should be raised.

Note

for loops expect that an IndexError will be raised for illegal indexes to allow proper detection of the end of the sequence.

Note

When subscripting a class, the special class method __class_getitem__() may be called instead of __getitem__(). See __class_getitem__ versus __getitem__ for more details.

object.__setitem__(selfkeyvalue)

Called to implement assignment to self[key]. Same note as for __getitem__(). This should only be implemented for mappings if the objects support changes to the values for keys, or if new keys can be added, or for sequences if elements can be replaced. The same exceptions should be raised for improper key values as for the __getitem__() method.

object.__delitem__(selfkey)

Called to implement deletion of self[key]. Same note as for __getitem__(). This should only be implemented for mappings if the objects support removal of keys, or for sequences if elements can be removed from the sequence. The same exceptions should be raised for improper key values as for the __getitem__() method.

object.__missing__(selfkey)

Called by dict.__getitem__() to implement self[key] for dict subclasses when key is not in the dictionary.

object.__iter__(self)

This method is called when an iterator is required for a container. This method should return a new iterator object that can iterate over all the objects in the container. For mappings, it should iterate over the keys of the container.

object.__reversed__(self)

Called (if present) by the reversed() built-in to implement reverse iteration. It should return a new iterator object that iterates over all the objects in the container in reverse order.

If the __reversed__() method is not provided, the reversed() built-in will fall back to using the sequence protocol (__len__() and __getitem__()). Objects that support the sequence protocol should only provide __reversed__() if they can provide an implementation that is more efficient than the one provided by reversed().

The membership test operators (in and not in) are normally implemented as an iteration through a container. However, container objects can supply the following special method with a more efficient implementation, which also does not require the object be iterable.

object.__contains__(selfitem)

Called to implement membership test operators. Should return true if item is in self, false otherwise. For mapping objects, this should consider the keys of the mapping rather than the values or the key-item pairs.

For objects that don’t define __contains__(), the membership test first tries iteration via __iter__(), then the old sequence iteration protocol via __getitem__(), see this section in the language reference.

3.3.8. Emulating numeric types

The following methods can be defined to emulate numeric objects. Methods corresponding to operations that are not supported by the particular kind of number implemented (e.g., bitwise operations for non-integral numbers) should be left undefined.

object.__add__(selfother)
object.__sub__(selfother)
object.__mul__(selfother)
object.__matmul__(selfother)
object.__truediv__(selfother)
object.__floordiv__(selfother)
object.__mod__(selfother)
object.__divmod__(selfother)
object.__pow__(selfother[modulo])
object.__lshift__(selfother)
object.__rshift__(selfother)
object.__and__(selfother)
object.__xor__(selfother)
object.__or__(selfother)

These methods are called to implement the binary arithmetic operations (+-*@///%divmod()pow()**<<>>&^|). For instance, to evaluate the expression x + y, where x is an instance of a class that has an __add__() method, x.__add__(y) is called. The __divmod__() method should be the equivalent to using __floordiv__() and __mod__(); it should not be related to __truediv__(). Note that __pow__() should be defined to accept an optional third argument if the ternary version of the built-in pow() function is to be supported.

If one of those methods does not support the operation with the supplied arguments, it should return NotImplemented.

object.__radd__(selfother)
object.__rsub__(selfother)
object.__rmul__(selfother)
object.__rmatmul__(selfother)
object.__rtruediv__(selfother)
object.__rfloordiv__(selfother)
object.__rmod__(selfother)
object.__rdivmod__(selfother)
object.__rpow__(selfother[modulo])
object.__rlshift__(selfother)
object.__rrshift__(selfother)
object.__rand__(selfother)
object.__rxor__(selfother)
object.__ror__(selfother)

These methods are called to implement the binary arithmetic operations (+-*@///%divmod()pow()**<<>>&^|) with reflected (swapped) operands. These functions are only called if the left operand does not support the corresponding operation 3 and the operands are of different types. 4 For instance, to evaluate the expression x - y, where y is an instance of a class that has an __rsub__() method, y.__rsub__(x) is called if x.__sub__(y) returns NotImplemented.

Note that ternary pow() will not try calling __rpow__() (the coercion rules would become too complicated).

Note

If the right operand’s type is a subclass of the left operand’s type and that subclass provides a different implementation of the reflected method for the operation, this method will be called before the left operand’s non-reflected method. This behavior allows subclasses to override their ancestors’ operations.

object.__iadd__(selfother)
object.__isub__(selfother)
object.__imul__(selfother)
object.__imatmul__(selfother)
object.__itruediv__(selfother)
object.__ifloordiv__(selfother)
object.__imod__(selfother)
object.__ipow__(selfother[modulo])
object.__ilshift__(selfother)
object.__irshift__(selfother)
object.__iand__(selfother)
object.__ixor__(selfother)
object.__ior__(selfother)

These methods are called to implement the augmented arithmetic assignments (+=-=*=@=/=//=%=**=<<=>>=&=^=|=). These methods should attempt to do the operation in-place (modifying self) and return the result (which could be, but does not have to be, self). If a specific method is not defined, the augmented assignment falls back to the normal methods. For instance, if x is an instance of a class with an __iadd__() method, x += y is equivalent to x = x.__iadd__(y) . Otherwise, x.__add__(y) and y.__radd__(x) are considered, as with the evaluation of x + y. In certain situations, augmented assignment can result in unexpected errors (see Why does a_tuple[i] += [‘item’] raise an exception when the addition works?), but this behavior is in fact part of the data model.

object.__neg__(self)
object.__pos__(self)
object.__abs__(self)
object.__invert__(self)

Called to implement the unary arithmetic operations (-+abs() and ~).

object.__complex__(self)
object.__int__(self)
object.__float__(self)

Called to implement the built-in functions complex()int() and float(). Should return a value of the appropriate type.

object.__index__(self)

Called to implement operator.index(), and whenever Python needs to losslessly convert the numeric object to an integer object (such as in slicing, or in the built-in bin()hex() and oct() functions). Presence of this method indicates that the numeric object is an integer type. Must return an integer.

If __int__()__float__() and __complex__() are not defined then corresponding built-in functions int()float() and complex() fall back to __index__().

object.__round__(self[ndigits])
object.__trunc__(self)
object.__floor__(self)
object.__ceil__(self)

Called to implement the built-in function round() and math functions trunc()floor() and ceil(). Unless ndigits is passed to __round__() all these methods should return the value of the object truncated to an Integral (typically an int).

The built-in function int() falls back to __trunc__() if neither __int__() nor __index__() is defined.

3.3.9. With Statement Context Managers

context manager is an object that defines the runtime context to be established when executing a with statement. The context manager handles the entry into, and the exit from, the desired runtime context for the execution of the block of code. Context managers are normally invoked using the with statement (described in section The with statement), but can also be used by directly invoking their methods.

Typical uses of context managers include saving and restoring various kinds of global state, locking and unlocking resources, closing opened files, etc.

For more information on context managers, see Context Manager Types.

object.__enter__(self)

Enter the runtime context related to this object. The with statement will bind this method’s return value to the target(s) specified in the as clause of the statement, if any.

object.__exit__(selfexc_typeexc_valuetraceback)

Exit the runtime context related to this object. The parameters describe the exception that caused the context to be exited. If the context was exited without an exception, all three arguments will be None.

If an exception is supplied, and the method wishes to suppress the exception (i.e., prevent it from being propagated), it should return a true value. Otherwise, the exception will be processed normally upon exit from this method.

Note that __exit__() methods should not reraise the passed-in exception; this is the caller’s responsibility.

See also

PEP 343 – The “with” statement

The specification, background, and examples for the Python with statement.

3.3.10. Customizing positional arguments in class pattern matching

When using a class name in a pattern, positional arguments in the pattern are not allowed by default, i.e. case MyClass(x, y) is typically invalid without special support in MyClass. To be able to use that kind of patterns, the class needs to define a __match_args__ attribute.

object.__match_args__

This class variable can be assigned a tuple of strings. When this class is used in a class pattern with positional arguments, each positional argument will be converted into a keyword argument, using the corresponding value in __match_args__ as the keyword. The absence of this attribute is equivalent to setting it to ().

For example, if MyClass.__match_args__ is ("left", "center", "right") that means that case MyClass(x, y) is equivalent to case MyClass(left=x, center=y). Note that the number of arguments in the pattern must be smaller than or equal to the number of elements in __match_args__; if it is larger, the pattern match attempt will raise a TypeError.

New in version 3.10.

See also

PEP 634 – Structural Pattern Matching

The specification for the Python match statement.

3.3.11. Special method lookup

For custom classes, implicit invocations of special methods are only guaranteed to work correctly if defined on an object’s type, not in the object’s instance dictionary. That behaviour is the reason why the following code raises an exception:

>>>

>>> class C:
...     pass
...
>>> c = C()
>>> c.__len__ = lambda: 5
>>> len(c)
Traceback (most recent call last):
  File "", line 1, in 
TypeError: object of type 'C' has no len()

The rationale behind this behaviour lies with a number of special methods such as __hash__() and __repr__() that are implemented by all objects, including type objects. If the implicit lookup of these methods used the conventional lookup process, they would fail when invoked on the type object itself:

>>>

>>> 1 .__hash__() == hash(1)
True
>>> int.__hash__() == hash(int)
Traceback (most recent call last):
  File "", line 1, in 
TypeError: descriptor '__hash__' of 'int' object needs an argument

Incorrectly attempting to invoke an unbound method of a class in this way is sometimes referred to as ‘metaclass confusion’, and is avoided by bypassing the instance when looking up special methods:

>>>

>>> type(1).__hash__(1) == hash(1)
True
>>> type(int).__hash__(int) == hash(int)
True

In addition to bypassing any instance attributes in the interest of correctness, implicit special method lookup generally also bypasses the __getattribute__() method even of the object’s metaclass:

>>>

>>> class Meta(type):
...     def __getattribute__(*args):
...         print("Metaclass getattribute invoked")
...         return type.__getattribute__(*args)
...
>>> class C(object, metaclass=Meta):
...     def __len__(self):
...         return 10
...     def __getattribute__(*args):
...         print("Class getattribute invoked")
...         return object.__getattribute__(*args)
...
>>> c = C()
>>> c.__len__()                 # Explicit lookup via instance
Class getattribute invoked
10
>>> type(c).__len__(c)          # Explicit lookup via type
Metaclass getattribute invoked
10
>>> len(c)                      # Implicit lookup
10

Bypassing the __getattribute__() machinery in this fashion provides significant scope for speed optimisations within the interpreter, at the cost of some flexibility in the handling of special methods (the special method must be set on the class object itself in order to be consistently invoked by the interpreter).

3.4. Coroutines

3.4.1. Awaitable Objects

An awaitable object generally implements an __await__() method. Coroutine objects returned from async def functions are awaitable.

Note

The generator iterator objects returned from generators decorated with types.coroutine() or asyncio.coroutine() are also awaitable, but they do not implement __await__().

object.__await__(self)

Must return an iterator. Should be used to implement awaitable objects. For instance, asyncio.Future implements this method to be compatible with the await expression.

New in version 3.5.

See also

PEP 492 for additional information about awaitable objects.

3.4.2. Coroutine Objects

Coroutine objects are awaitable objects. A coroutine’s execution can be controlled by calling __await__() and iterating over the result. When the coroutine has finished executing and returns, the iterator raises StopIteration, and the exception’s value attribute holds the return value. If the coroutine raises an exception, it is propagated by the iterator. Coroutines should not directly raise unhandled StopIteration exceptions.

Coroutines also have the methods listed below, which are analogous to those of generators (see Generator-iterator methods). However, unlike generators, coroutines do not directly support iteration.

Changed in version 3.5.2: It is a RuntimeError to await on a coroutine more than once.

coroutine.send(value)

Starts or resumes execution of the coroutine. If value is None, this is equivalent to advancing the iterator returned by __await__(). If value is not None, this method delegates to the send() method of the iterator that caused the coroutine to suspend. The result (return value, StopIteration, or other exception) is the same as when iterating over the __await__() return value, described above.

coroutine.throw(value)
coroutine.throw(type[value[traceback]])

Raises the specified exception in the coroutine. This method delegates to the throw() method of the iterator that caused the coroutine to suspend, if it has such a method. Otherwise, the exception is raised at the suspension point. The result (return value, StopIteration, or other exception) is the same as when iterating over the __await__() return value, described above. If the exception is not caught in the coroutine, it propagates back to the caller.

coroutine.close()

Causes the coroutine to clean itself up and exit. If the coroutine is suspended, this method first delegates to the close() method of the iterator that caused the coroutine to suspend, if it has such a method. Then it raises GeneratorExit at the suspension point, causing the coroutine to immediately clean itself up. Finally, the coroutine is marked as having finished executing, even if it was never started.

Coroutine objects are automatically closed using the above process when they are about to be destroyed.

3.4.3. Asynchronous Iterators

An asynchronous iterator can call asynchronous code in its __anext__ method.

Asynchronous iterators can be used in an async for statement.

object.__aiter__(self)

Must return an asynchronous iterator object.

object.__anext__(self)

Must return an awaitable resulting in a next value of the iterator. Should raise a StopAsyncIteration error when the iteration is over.

An example of an asynchronous iterable object:

class Reader:
    async def readline(self):
        ...

    def __aiter__(self):
        return self

    async def __anext__(self):
        val = await self.readline()
        if val == b'':
            raise StopAsyncIteration
        return val

New in version 3.5.

Changed in version 3.7: Prior to Python 3.7, __aiter__() could return an awaitable that would resolve to an asynchronous iterator.

Starting with Python 3.7, __aiter__() must return an asynchronous iterator object. Returning anything else will result in a TypeError error.

3.4.4. Asynchronous Context Managers

An asynchronous context manager is a context manager that is able to suspend execution in its __aenter__ and __aexit__ methods.

Asynchronous context managers can be used in an async with statement.

object.__aenter__(self)

Semantically similar to __enter__(), the only difference being that it must return an awaitable.

object.__aexit__(selfexc_typeexc_valuetraceback)

Semantically similar to __exit__(), the only difference being that it must return an awaitable.

An example of an asynchronous context manager class:

class AsyncContextManager:
    async def __aenter__(self):
        await log('entering context')

    async def __aexit__(self, exc_type, exc, tb):
        await log('exiting context')

New in version 3.5.

Footnotes

1

It is possible in some cases to change an object’s type, under certain controlled conditions. It generally isn’t a good idea though, since it can lead to some very strange behaviour if it is handled incorrectly.

2

The __hash__()__iter__()__reversed__(), and __contains__() methods have special handling for this; others will still raise a TypeError, but may do so by relying on the behavior that None is not callable.

3

“Does not support” here means that the class has no such method, or the method returns NotImplemented. Do not set the method to None if you want to force fallback to the right operand’s reflected method—that will instead have the opposite effect of explicitly blocking such fallback.

4

For operands of the same type, it is assumed that if the non-reflected method – such as __add__() – fails then the overall operation is not supported, which is why the reflected method is not called.

4. Execution model

4.1. Structure of a program

A Python program is constructed from code blocks. A block is a piece of Python program text that is executed as a unit. The following are blocks: a module, a function body, and a class definition. Each command typed interactively is a block. A script file (a file given as standard input to the interpreter or specified as a command line argument to the interpreter) is a code block. A script command (a command specified on the interpreter command line with the -c option) is a code block. A module run as a top level script (as module __main__) from the command line using a -m argument is also a code block. The string argument passed to the built-in functions eval() and exec() is a code block.

A code block is executed in an execution frame. A frame contains some administrative information (used for debugging) and determines where and how execution continues after the code block’s execution has completed.

4.2. Naming and binding

4.2.1. Binding of names

Names refer to objects. Names are introduced by name binding operations.

The following constructs bind names:

  • formal parameters to functions,

  • class definitions,

  • function definitions,

  • assignment expressions,

  • targets that are identifiers if occurring in an assignment:

    • for loop header,

    • after as in a with statement, except clause or in the as-pattern in structural pattern matching,

    • in a capture pattern in structural pattern matching

  • import statements.

The import statement of the form from ... import * binds all names defined in the imported module, except those beginning with an underscore. This form may only be used at the module level.

A target occurring in a del statement is also considered bound for this purpose (though the actual semantics are to unbind the name).

Each assignment or import statement occurs within a block defined by a class or function definition or at the module level (the top-level code block).

If a name is bound in a block, it is a local variable of that block, unless declared as nonlocal or global. If a name is bound at the module level, it is a global variable. (The variables of the module code block are local and global.) If a variable is used in a code block but not defined there, it is a free variable.

Each occurrence of a name in the program text refers to the binding of that name established by the following name resolution rules.

4.2.2. Resolution of names

scope defines the visibility of a name within a block. If a local variable is defined in a block, its scope includes that block. If the definition occurs in a function block, the scope extends to any blocks contained within the defining one, unless a contained block introduces a different binding for the name.

When a name is used in a code block, it is resolved using the nearest enclosing scope. The set of all such scopes visible to a code block is called the block’s environment.

When a name is not found at all, a NameError exception is raised. If the current scope is a function scope, and the name refers to a local variable that has not yet been bound to a value at the point where the name is used, an UnboundLocalError exception is raised. UnboundLocalError is a subclass of NameError.

If a name binding operation occurs anywhere within a code block, all uses of the name within the block are treated as references to the current block. This can lead to errors when a name is used within a block before it is bound. This rule is subtle. Python lacks declarations and allows name binding operations to occur anywhere within a code block. The local variables of a code block can be determined by scanning the entire text of the block for name binding operations.

If the global statement occurs within a block, all uses of the names specified in the statement refer to the bindings of those names in the top-level namespace. Names are resolved in the top-level namespace by searching the global namespace, i.e. the namespace of the module containing the code block, and the builtins namespace, the namespace of the module builtins. The global namespace is searched first. If the names are not found there, the builtins namespace is searched. The global statement must precede all uses of the listed names.

The global statement has the same scope as a name binding operation in the same block. If the nearest enclosing scope for a free variable contains a global statement, the free variable is treated as a global.

The nonlocal statement causes corresponding names to refer to previously bound variables in the nearest enclosing function scope. SyntaxError is raised at compile time if the given name does not exist in any enclosing function scope.

The namespace for a module is automatically created the first time a module is imported. The main module for a script is always called __main__.

Class definition blocks and arguments to exec() and eval() are special in the context of name resolution. A class definition is an executable statement that may use and define names. These references follow the normal rules for name resolution with an exception that unbound local variables are looked up in the global namespace. The namespace of the class definition becomes the attribute dictionary of the class. The scope of names defined in a class block is limited to the class block; it does not extend to the code blocks of methods – this includes comprehensions and generator expressions since they are implemented using a function scope. This means that the following will fail:

class A:
    a = 42
    b = list(a + i for i in range(10))

4.2.3. Builtins and restricted execution

CPython implementation detail: Users should not touch __builtins__; it is strictly an implementation detail. Users wanting to override values in the builtins namespace should import the builtins module and modify its attributes appropriately.

The builtins namespace associated with the execution of a code block is actually found by looking up the name __builtins__ in its global namespace; this should be a dictionary or a module (in the latter case the module’s dictionary is used). By default, when in the __main__ module, __builtins__ is the built-in module builtins; when in any other module, __builtins__ is an alias for the dictionary of the builtins module itself.

4.2.4. Interaction with dynamic features

Name resolution of free variables occurs at runtime, not at compile time. This means that the following code will print 42:

i = 10
def f():
    print(i)
i = 42
f()

The eval() and exec() functions do not have access to the full environment for resolving names. Names may be resolved in the local and global namespaces of the caller. Free variables are not resolved in the nearest enclosing namespace, but in the global namespace. 1 The exec() and eval() functions have optional arguments to override the global and local namespace. If only one namespace is specified, it is used for both.

4.3. Exceptions

 

Exceptions are a means of breaking out of the normal flow of control of a code block in order to handle errors or other exceptional conditions. An exception is raised at the point where the error is detected; it may be handled by the surrounding code block or by any code block that directly or indirectly invoked the code block where the error occurred.

The Python interpreter raises an exception when it detects a run-time error (such as division by zero). A Python program can also explicitly raise an exception with the raise statement. Exception handlers are specified with the try … except statement. The finally clause of such a statement can be used to specify cleanup code which does not handle the exception, but is executed whether an exception occurred or not in the preceding code.

Python uses the “termination” model of error handling: an exception handler can find out what happened and continue execution at an outer level, but it cannot repair the cause of the error and retry the failing operation (except by re-entering the offending piece of code from the top).

When an exception is not handled at all, the interpreter terminates execution of the program, or returns to its interactive main loop. In either case, it prints a stack traceback, except when the exception is SystemExit.

Exceptions are identified by class instances. The except clause is selected depending on the class of the instance: it must reference the class of the instance or a non-virtual base class thereof. The instance can be received by the handler and can carry additional information about the exceptional condition.

Note

Exception messages are not part of the Python API. Their contents may change from one version of Python to the next without warning and should not be relied on by code which will run under multiple versions of the interpreter.

See also the description of the try statement in section The try statement and raise statement in section The raise statement.

Footnotes

1

This limitation occurs because the code that is executed by these operations is not available at the time the module is compiled.

5. The import system

Python code in one module gains access to the code in another module by the process of importing it. The import statement is the most common way of invoking the import machinery, but it is not the only way. Functions such as importlib.import_module() and built-in __import__() can also be used to invoke the import machinery.

The import statement combines two operations; it searches for the named module, then it binds the results of that search to a name in the local scope. The search operation of the import statement is defined as a call to the __import__() function, with the appropriate arguments. The return value of __import__() is used to perform the name binding operation of the import statement. See the import statement for the exact details of that name binding operation.

A direct call to __import__() performs only the module search and, if found, the module creation operation. While certain side-effects may occur, such as the importing of parent packages, and the updating of various caches (including sys.modules), only the import statement performs a name binding operation.

When an import statement is executed, the standard builtin __import__() function is called. Other mechanisms for invoking the import system (such as importlib.import_module()) may choose to bypass __import__() and use their own solutions to implement import semantics.

When a module is first imported, Python searches for the module and if found, it creates a module object 1, initializing it. If the named module cannot be found, a ModuleNotFoundError is raised. Python implements various strategies to search for the named module when the import machinery is invoked. These strategies can be modified and extended by using various hooks described in the sections below.

Changed in version 3.3: The import system has been updated to fully implement the second phase of PEP 302. There is no longer any implicit import machinery – the full import system is exposed through sys.meta_path. In addition, native namespace package support has been implemented (see PEP 420).

5.1. importlib

The importlib module provides a rich API for interacting with the import system. For example importlib.import_module() provides a recommended, simpler API than built-in __import__() for invoking the import machinery. Refer to the importlib library documentation for additional detail.

5.2. Packages

Python has only one type of module object, and all modules are of this type, regardless of whether the module is implemented in Python, C, or something else. To help organize modules and provide a naming hierarchy, Python has a concept of packages.

You can think of packages as the directories on a file system and modules as files within directories, but don’t take this analogy too literally since packages and modules need not originate from the file system. For the purposes of this documentation, we’ll use this convenient analogy of directories and files. Like file system directories, packages are organized hierarchically, and packages may themselves contain subpackages, as well as regular modules.

It’s important to keep in mind that all packages are modules, but not all modules are packages. Or put another way, packages are just a special kind of module. Specifically, any module that contains a __path__ attribute is considered a package.

All modules have a name. Subpackage names are separated from their parent package name by a dot, akin to Python’s standard attribute access syntax. Thus you might have a package called email, which in turn has a subpackage called email.mime and a module within that subpackage called email.mime.text.

5.2.1. Regular packages

Python defines two types of packages, regular packages and namespace packages. Regular packages are traditional packages as they existed in Python 3.2 and earlier. A regular package is typically implemented as a directory containing an __init__.py file. When a regular package is imported, this __init__.py file is implicitly executed, and the objects it defines are bound to names in the package’s namespace. The __init__.py file can contain the same Python code that any other module can contain, and Python will add some additional attributes to the module when it is imported.

For example, the following file system layout defines a top level parent package with three subpackages:

parent/
    __init__.py
    one/
        __init__.py
    two/
        __init__.py
    three/
        __init__.py

Importing parent.one will implicitly execute parent/__init__.py and parent/one/__init__.py. Subsequent imports of parent.two or parent.three will execute parent/two/__init__.py and parent/three/__init__.py respectively.

5.2.2. Namespace packages

A namespace package is a composite of various portions, where each portion contributes a subpackage to the parent package. Portions may reside in different locations on the file system. Portions may also be found in zip files, on the network, or anywhere else that Python searches during import. Namespace packages may or may not correspond directly to objects on the file system; they may be virtual modules that have no concrete representation.

Namespace packages do not use an ordinary list for their __path__ attribute. They instead use a custom iterable type which will automatically perform a new search for package portions on the next import attempt within that package if the path of their parent package (or sys.path for a top level package) changes.

With namespace packages, there is no parent/__init__.py file. In fact, there may be multiple parent directories found during import search, where each one is provided by a different portion. Thus parent/one may not be physically located next to parent/two. In this case, Python will create a namespace package for the top-level parent package whenever it or one of its subpackages is imported.

See also PEP 420 for the namespace package specification.

5.3. Searching

To begin the search, Python needs the fully qualified name of the module (or package, but for the purposes of this discussion, the difference is immaterial) being imported. This name may come from various arguments to the import statement, or from the parameters to the importlib.import_module() or __import__() functions.

This name will be used in various phases of the import search, and it may be the dotted path to a submodule, e.g. foo.bar.baz. In this case, Python first tries to import foo, then foo.bar, and finally foo.bar.baz. If any of the intermediate imports fail, a ModuleNotFoundError is raised.

5.3.1. The module cache

The first place checked during import search is sys.modules. This mapping serves as a cache of all modules that have been previously imported, including the intermediate paths. So if foo.bar.baz was previously imported, sys.modules will contain entries for foofoo.bar, and foo.bar.baz. Each key will have as its value the corresponding module object.

During import, the module name is looked up in sys.modules and if present, the associated value is the module satisfying the import, and the process completes. However, if the value is None, then a ModuleNotFoundError is raised. If the module name is missing, Python will continue searching for the module.

sys.modules is writable. Deleting a key may not destroy the associated module (as other modules may hold references to it), but it will invalidate the cache entry for the named module, causing Python to search anew for the named module upon its next import. The key can also be assigned to None, forcing the next import of the module to result in a ModuleNotFoundError.

Beware though, as if you keep a reference to the module object, invalidate its cache entry in sys.modules, and then re-import the named module, the two module objects will not be the same. By contrast, importlib.reload() will reuse the same module object, and simply reinitialise the module contents by rerunning the module’s code.

5.3.2. Finders and loaders

If the named module is not found in sys.modules, then Python’s import protocol is invoked to find and load the module. This protocol consists of two conceptual objects, finders and loaders. A finder’s job is to determine whether it can find the named module using whatever strategy it knows about. Objects that implement both of these interfaces are referred to as importers – they return themselves when they find that they can load the requested module.

Python includes a number of default finders and importers. The first one knows how to locate built-in modules, and the second knows how to locate frozen modules. A third default finder searches an import path for modules. The import path is a list of locations that may name file system paths or zip files. It can also be extended to search for any locatable resource, such as those identified by URLs.

The import machinery is extensible, so new finders can be added to extend the range and scope of module searching.

Finders do not actually load modules. If they can find the named module, they return a module spec, an encapsulation of the module’s import-related information, which the import machinery then uses when loading the module.

The following sections describe the protocol for finders and loaders in more detail, including how you can create and register new ones to extend the import machinery.

Changed in version 3.4: In previous versions of Python, finders returned loaders directly, whereas now they return module specs which contain loaders. Loaders are still used during import but have fewer responsibilities.

5.3.3. Import hooks

The import machinery is designed to be extensible; the primary mechanism for this are the import hooks. There are two types of import hooks: meta hooks and import path hooks.

Meta hooks are called at the start of import processing, before any other import processing has occurred, other than sys.modules cache look up. This allows meta hooks to override sys.path processing, frozen modules, or even built-in modules. Meta hooks are registered by adding new finder objects to sys.meta_path, as described below.

Import path hooks are called as part of sys.path (or package.__path__) processing, at the point where their associated path item is encountered. Import path hooks are registered by adding new callables to sys.path_hooks as described below.

5.3.4. The meta path

When the named module is not found in sys.modules, Python next searches sys.meta_path, which contains a list of meta path finder objects. These finders are queried in order to see if they know how to handle the named module. Meta path finders must implement a method called find_spec() which takes three arguments: a name, an import path, and (optionally) a target module. The meta path finder can use any strategy it wants to determine whether it can handle the named module or not.

If the meta path finder knows how to handle the named module, it returns a spec object. If it cannot handle the named module, it returns None. If sys.meta_path processing reaches the end of its list without returning a spec, then a ModuleNotFoundError is raised. Any other exceptions raised are simply propagated up, aborting the import process.

The find_spec() method of meta path finders is called with two or three arguments. The first is the fully qualified name of the module being imported, for example foo.bar.baz. The second argument is the path entries to use for the module search. For top-level modules, the second argument is None, but for submodules or subpackages, the second argument is the value of the parent package’s __path__ attribute. If the appropriate __path__ attribute cannot be accessed, a ModuleNotFoundError is raised. The third argument is an existing module object that will be the target of loading later. The import system passes in a target module only during reload.

The meta path may be traversed multiple times for a single import request. For example, assuming none of the modules involved has already been cached, importing foo.bar.baz will first perform a top level import, calling mpf.find_spec("foo", None, None) on each meta path finder (mpf). After foo has been imported, foo.bar will be imported by traversing the meta path a second time, calling mpf.find_spec("foo.bar", foo.__path__, None). Once foo.bar has been imported, the final traversal will call mpf.find_spec("foo.bar.baz", foo.bar.__path__, None).

Some meta path finders only support top level imports. These importers will always return None when anything other than None is passed as the second argument.

Python’s default sys.meta_path has three meta path finders, one that knows how to import built-in modules, one that knows how to import frozen modules, and one that knows how to import modules from an import path (i.e. the path based finder).

Changed in version 3.4: The find_spec() method of meta path finders replaced find_module(), which is now deprecated. While it will continue to work without change, the import machinery will try it only if the finder does not implement find_spec().

Changed in version 3.10: Use of find_module() by the import system now raises ImportWarning.

5.4. Loading

If and when a module spec is found, the import machinery will use it (and the loader it contains) when loading the module. Here is an approximation of what happens during the loading portion of import:

module = None
if spec.loader is not None and hasattr(spec.loader, 'create_module'):
    # It is assumed 'exec_module' will also be defined on the loader.
    module = spec.loader.create_module(spec)
if module is None:
    module = ModuleType(spec.name)
# The import-related module attributes get set here:
_init_module_attrs(spec, module)

if spec.loader is None:
    # unsupported
    raise ImportError
if spec.origin is None and spec.submodule_search_locations is not None:
    # namespace package
    sys.modules[spec.name] = module
elif not hasattr(spec.loader, 'exec_module'):
    module = spec.loader.load_module(spec.name)
    # Set __loader__ and __package__ if missing.
else:
    sys.modules[spec.name] = module
    try:
        spec.loader.exec_module(module)
    except BaseException:
        try:
            del sys.modules[spec.name]
        except KeyError:
            pass
        raise
return sys.modules[spec.name]

Note the following details:

  • If there is an existing module object with the given name in sys.modules, import will have already returned it.

  • The module will exist in sys.modules before the loader executes the module code. This is crucial because the module code may (directly or indirectly) import itself; adding it to sys.modules beforehand prevents unbounded recursion in the worst case and multiple loading in the best.

  • If loading fails, the failing module – and only the failing module – gets removed from sys.modules. Any module already in the sys.modules cache, and any module that was successfully loaded as a side-effect, must remain in the cache. This contrasts with reloading where even the failing module is left in sys.modules.

  • After the module is created but before execution, the import machinery sets the import-related module attributes (“_init_module_attrs” in the pseudo-code example above), as summarized in a later section.

  • Module execution is the key moment of loading in which the module’s namespace gets populated. Execution is entirely delegated to the loader, which gets to decide what gets populated and how.

  • The module created during loading and passed to exec_module() may not be the one returned at the end of import 2.

Changed in version 3.4: The import system has taken over the boilerplate responsibilities of loaders. These were previously performed by the importlib.abc.Loader.load_module() method.

5.4.1. Loaders

Module loaders provide the critical function of loading: module execution. The import machinery calls the importlib.abc.Loader.exec_module() method with a single argument, the module object to execute. Any value returned from exec_module() is ignored.

Loaders must satisfy the following requirements:

  • If the module is a Python module (as opposed to a built-in module or a dynamically loaded extension), the loader should execute the module’s code in the module’s global name space (module.__dict__).

  • If the loader cannot execute the module, it should raise an ImportError, although any other exception raised during exec_module() will be propagated.

In many cases, the finder and loader can be the same object; in such cases the find_spec() method would just return a spec with the loader set to self.

Module loaders may opt in to creating the module object during loading by implementing a create_module() method. It takes one argument, the module spec, and returns the new module object to use during loading. create_module() does not need to set any attributes on the module object. If the method returns None, the import machinery will create the new module itself.

New in version 3.4: The create_module() method of loaders.

Changed in version 3.4: The load_module() method was replaced by exec_module() and the import machinery assumed all the boilerplate responsibilities of loading.

For compatibility with existing loaders, the import machinery will use the load_module() method of loaders if it exists and the loader does not also implement exec_module(). However, load_module() has been deprecated and loaders should implement exec_module() instead.

The load_module() method must implement all the boilerplate loading functionality described above in addition to executing the module. All the same constraints apply, with some additional clarification:

  • If there is an existing module object with the given name in sys.modules, the loader must use that existing module. (Otherwise, importlib.reload() will not work correctly.) If the named module does not exist in sys.modules, the loader must create a new module object and add it to sys.modules.

  • The module must exist in sys.modules before the loader executes the module code, to prevent unbounded recursion or multiple loading.

  • If loading fails, the loader must remove any modules it has inserted into sys.modules, but it must remove only the failing module(s), and only if the loader itself has loaded the module(s) explicitly.

Changed in version 3.5: DeprecationWarning is raised when exec_module() is defined but create_module() is not.

Changed in version 3.6: An ImportError is raised when exec_module() is defined but create_module() is not.

Changed in version 3.10: Use of load_module() will raise ImportWarning.

5.4.2. Submodules

When a submodule is loaded using any mechanism (e.g. importlib APIs, the import or import-from statements, or built-in __import__()) a binding is placed in the parent module’s namespace to the submodule object. For example, if package spam has a submodule foo, after importing spam.foospam will have an attribute foo which is bound to the submodule. Let’s say you have the following directory structure:

spam/
    __init__.py
    foo.py

and spam/__init__.py has the following line in it:

from .foo import Foo

then executing the following puts name bindings for foo and Foo in the spam module:

>>>

>>> import spam
>>> spam.foo
<module 'spam.foo' from '/tmp/imports/spam/foo.py'>
>>> spam.Foo
<class 'spam.foo.Foo'>

Given Python’s familiar name binding rules this might seem surprising, but it’s actually a fundamental feature of the import system. The invariant holding is that if you have sys.modules['spam'] and sys.modules['spam.foo'] (as you would after the above import), the latter must appear as the foo attribute of the former.

5.4.3. Module spec

The import machinery uses a variety of information about each module during import, especially before loading. Most of the information is common to all modules. The purpose of a module’s spec is to encapsulate this import-related information on a per-module basis.

Using a spec during import allows state to be transferred between import system components, e.g. between the finder that creates the module spec and the loader that executes it. Most importantly, it allows the import machinery to perform the boilerplate operations of loading, whereas without a module spec the loader had that responsibility.

The module’s spec is exposed as the __spec__ attribute on a module object. See ModuleSpec for details on the contents of the module spec.

New in version 3.4.

5.4.5. module.__path__

By definition, if a module has a __path__ attribute, it is a package.

A package’s __path__ attribute is used during imports of its subpackages. Within the import machinery, it functions much the same as sys.path, i.e. providing a list of locations to search for modules during import. However, __path__ is typically much more constrained than sys.path.

__path__ must be an iterable of strings, but it may be empty. The same rules used for sys.path also apply to a package’s __path__, and sys.path_hooks (described below) are consulted when traversing a package’s __path__.

A package’s __init__.py file may set or alter the package’s __path__ attribute, and this was typically the way namespace packages were implemented prior to PEP 420. With the adoption of PEP 420, namespace packages no longer need to supply __init__.py files containing only __path__ manipulation code; the import machinery automatically sets __path__ correctly for the namespace package.

5.4.6. Module reprs

By default, all modules have a usable repr, however depending on the attributes set above, and in the module’s spec, you can more explicitly control the repr of module objects.

If the module has a spec (__spec__), the import machinery will try to generate a repr from it. If that fails or there is no spec, the import system will craft a default repr using whatever information is available on the module. It will try to use the module.__name__module.__file__, and module.__loader__ as input into the repr, with defaults for whatever information is missing.

Here are the exact rules used:

  • If the module has a __spec__ attribute, the information in the spec is used to generate the repr. The “name”, “loader”, “origin”, and “has_location” attributes are consulted.

  • If the module has a __file__ attribute, this is used as part of the module’s repr.

  • If the module has no __file__ but does have a __loader__ that is not None, then the loader’s repr is used as part of the module’s repr.

  • Otherwise, just use the module’s __name__ in the repr.

Changed in version 3.4: Use of loader.module_repr() has been deprecated and the module spec is now used by the import machinery to generate a module repr.

For backward compatibility with Python 3.3, the module repr will be generated by calling the loader’s module_repr() method, if defined, before trying either approach described above. However, the method is deprecated.

Changed in version 3.10: Calling module_repr() now occurs after trying to use a module’s __spec__ attribute but before falling back on __file__. Use of module_repr() is slated to stop in Python 3.12.

5.4.7. Cached bytecode invalidation

Before Python loads cached bytecode from a .pyc file, it checks whether the cache is up-to-date with the source .py file. By default, Python does this by storing the source’s last-modified timestamp and size in the cache file when writing it. At runtime, the import system then validates the cache file by checking the stored metadata in the cache file against the source’s metadata.

Python also supports “hash-based” cache files, which store a hash of the source file’s contents rather than its metadata. There are two variants of hash-based .pyc files: checked and unchecked. For checked hash-based .pyc files, Python validates the cache file by hashing the source file and comparing the resulting hash with the hash in the cache file. If a checked hash-based cache file is found to be invalid, Python regenerates it and writes a new checked hash-based cache file. For unchecked hash-based .pyc files, Python simply assumes the cache file is valid if it exists. Hash-based .pyc files validation behavior may be overridden with the --check-hash-based-pycs flag.

Changed in version 3.7: Added hash-based .pyc files. Previously, Python only supported timestamp-based invalidation of bytecode caches.

5.5. The Path Based Finder

As mentioned previously, Python comes with several default meta path finders. One of these, called the path based finder (PathFinder), searches an import path, which contains a list of path entries. Each path entry names a location to search for modules.

The path based finder itself doesn’t know how to import anything. Instead, it traverses the individual path entries, associating each of them with a path entry finder that knows how to handle that particular kind of path.

The default set of path entry finders implement all the semantics for finding modules on the file system, handling special file types such as Python source code (.py files), Python byte code (.pyc files) and shared libraries (e.g. .so files). When supported by the zipimport module in the standard library, the default path entry finders also handle loading all of these file types (other than shared libraries) from zipfiles.

Path entries need not be limited to file system locations. They can refer to URLs, database queries, or any other location that can be specified as a string.

The path based finder provides additional hooks and protocols so that you can extend and customize the types of searchable path entries. For example, if you wanted to support path entries as network URLs, you could write a hook that implements HTTP semantics to find modules on the web. This hook (a callable) would return a path entry finder supporting the protocol described below, which was then used to get a loader for the module from the web.

A word of warning: this section and the previous both use the term finder, distinguishing between them by using the terms meta path finder and path entry finder. These two types of finders are very similar, support similar protocols, and function in similar ways during the import process, but it’s important to keep in mind that they are subtly different. In particular, meta path finders operate at the beginning of the import process, as keyed off the sys.meta_path traversal.

By contrast, path entry finders are in a sense an implementation detail of the path based finder, and in fact, if the path based finder were to be removed from sys.meta_path, none of the path entry finder semantics would be invoked.

5.5.1. Path entry finders

The path based finder is responsible for finding and loading Python modules and packages whose location is specified with a string path entry. Most path entries name locations in the file system, but they need not be limited to this.

As a meta path finder, the path based finder implements the find_spec() protocol previously described, however it exposes additional hooks that can be used to customize how modules are found and loaded from the import path.

Three variables are used by the path based findersys.pathsys.path_hooks and sys.path_importer_cache. The __path__ attributes on package objects are also used. These provide additional ways that the import machinery can be customized.

sys.path contains a list of strings providing search locations for modules and packages. It is initialized from the PYTHONPATH environment variable and various other installation- and implementation-specific defaults. Entries in sys.path can name directories on the file system, zip files, and potentially other “locations” (see the site module) that should be searched for modules, such as URLs, or database queries. Only strings and bytes should be present on sys.path; all other data types are ignored. The encoding of bytes entries is determined by the individual path entry finders.

The path based finder is a meta path finder, so the import machinery begins the import path search by calling the path based finder’s find_spec() method as described previously. When the path argument to find_spec() is given, it will be a list of string paths to traverse – typically a package’s __path__ attribute for an import within that package. If the path argument is None, this indicates a top level import and sys.path is used.

The path based finder iterates over every entry in the search path, and for each of these, looks for an appropriate path entry finder (PathEntryFinder) for the path entry. Because this can be an expensive operation (e.g. there may be stat() call overheads for this search), the path based finder maintains a cache mapping path entries to path entry finders. This cache is maintained in sys.path_importer_cache (despite the name, this cache actually stores finder objects rather than being limited to importer objects). In this way, the expensive search for a particular path entry location’s path entry finder need only be done once. User code is free to remove cache entries from sys.path_importer_cache forcing the path based finder to perform the path entry search again 3.

If the path entry is not present in the cache, the path based finder iterates over every callable in sys.path_hooks. Each of the path entry hooks in this list is called with a single argument, the path entry to be searched. This callable may either return a path entry finder that can handle the path entry, or it may raise ImportError. An ImportError is used by the path based finder to signal that the hook cannot find a path entry finder for that path entry. The exception is ignored and import path iteration continues. The hook should expect either a string or bytes object; the encoding of bytes objects is up to the hook (e.g. it may be a file system encoding, UTF-8, or something else), and if the hook cannot decode the argument, it should raise ImportError.

If sys.path_hooks iteration ends with no path entry finder being returned, then the path based finder’s find_spec() method will store None in sys.path_importer_cache (to indicate that there is no finder for this path entry) and return None, indicating that this meta path finder could not find the module.

If a path entry finder is returned by one of the path entry hook callables on sys.path_hooks, then the following protocol is used to ask the finder for a module spec, which is then used when loading the module.

The current working directory – denoted by an empty string – is handled slightly differently from other entries on sys.path. First, if the current working directory is found to not exist, no value is stored in sys.path_importer_cache. Second, the value for the current working directory is looked up fresh for each module lookup. Third, the path used for sys.path_importer_cache and returned by importlib.machinery.PathFinder.find_spec() will be the actual current working directory and not the empty string.

5.5.2. Path entry finder protocol

In order to support imports of modules and initialized packages and also to contribute portions to namespace packages, path entry finders must implement the find_spec() method.

find_spec() takes two arguments: the fully qualified name of the module being imported, and the (optional) target module. find_spec() returns a fully populated spec for the module. This spec will always have “loader” set (with one exception).

To indicate to the import machinery that the spec represents a namespace portion, the path entry finder sets “submodule_search_locations” to a list containing the portion.

Changed in version 3.4: find_spec() replaced find_loader() and find_module(), both of which are now deprecated, but will be used if find_spec() is not defined.

Older path entry finders may implement one of these two deprecated methods instead of find_spec(). The methods are still respected for the sake of backward compatibility. However, if find_spec() is implemented on the path entry finder, the legacy methods are ignored.

find_loader() takes one argument, the fully qualified name of the module being imported. find_loader() returns a 2-tuple where the first item is the loader and the second item is a namespace portion.

For backwards compatibility with other implementations of the import protocol, many path entry finders also support the same, traditional find_module() method that meta path finders support. However path entry finder find_module() methods are never called with a path argument (they are expected to record the appropriate path information from the initial call to the path hook).

The find_module() method on path entry finders is deprecated, as it does not allow the path entry finder to contribute portions to namespace packages. If both find_loader() and find_module() exist on a path entry finder, the import system will always call find_loader() in preference to find_module().

Changed in version 3.10: Calls to find_module() and find_loader() by the import system will raise ImportWarning.

5.6. Replacing the standard import system

The most reliable mechanism for replacing the entire import system is to delete the default contents of sys.meta_path, replacing them entirely with a custom meta path hook.

If it is acceptable to only alter the behaviour of import statements without affecting other APIs that access the import system, then replacing the builtin __import__() function may be sufficient. This technique may also be employed at the module level to only alter the behaviour of import statements within that module.

To selectively prevent the import of some modules from a hook early on the meta path (rather than disabling the standard import system entirely), it is sufficient to raise ModuleNotFoundError directly from find_spec() instead of returning None. The latter indicates that the meta path search should continue, while raising an exception terminates it immediately.

5.7. Package Relative Imports

Relative imports use leading dots. A single leading dot indicates a relative import, starting with the current package. Two or more leading dots indicate a relative import to the parent(s) of the current package, one level per dot after the first. For example, given the following package layout:

package/
    __init__.py
    subpackage1/
        __init__.py
        moduleX.py
        moduleY.py
    subpackage2/
        __init__.py
        moduleZ.py
    moduleA.py

In either subpackage1/moduleX.py or subpackage1/__init__.py, the following are valid relative imports:

from .moduleY import spam
from .moduleY import spam as ham
from . import moduleY
from ..subpackage1 import moduleY
from ..subpackage2.moduleZ import eggs
from ..moduleA import foo

Absolute imports may use either the import <> or from <> import <> syntax, but relative imports may only use the second form; the reason for this is that:

import XXX.YYY.ZZZ

should expose XXX.YYY.ZZZ as a usable expression, but .moduleY is not a valid expression.

5.8. Special considerations for __main__

The __main__ module is a special case relative to Python’s import system. As noted elsewhere, the __main__ module is directly initialized at interpreter startup, much like sys and builtins. However, unlike those two, it doesn’t strictly qualify as a built-in module. This is because the manner in which __main__ is initialized depends on the flags and other options with which the interpreter is invoked.

5.8.1. __main__.__spec__

Depending on how __main__ is initialized, __main__.__spec__ gets set appropriately or to None.

When Python is started with the -m option, __spec__ is set to the module spec of the corresponding module or package. __spec__ is also populated when the __main__ module is loaded as part of executing a directory, zipfile or other sys.path entry.

In the remaining cases __main__.__spec__ is set to None, as the code used to populate the __main__ does not correspond directly with an importable module:

  • interactive prompt

  • -c option

  • running from stdin

  • running directly from a source or bytecode file

Note that __main__.__spec__ is always None in the last case, even if the file could technically be imported directly as a module instead. Use the -m switch if valid module metadata is desired in __main__.

Note also that even when __main__ corresponds with an importable module and __main__.__spec__ is set accordingly, they’re still considered distinct modules. This is due to the fact that blocks guarded by if __name__ == "__main__": checks only execute when the module is used to populate the __main__ namespace, and not during normal import.

5.9. Open issues

XXX It would be really nice to have a diagram.

XXX * (import_machinery.rst) how about a section devoted just to the attributes of modules and packages, perhaps expanding upon or supplanting the related entries in the data model reference page?

XXX runpy, pkgutil, et al in the library manual should all get “See Also” links at the top pointing to the new import system section.

XXX Add more explanation regarding the different ways in which __main__ is initialized?

XXX Add more info on __main__ quirks/pitfalls (i.e. copy from PEP 395).

5.10. References

The import machinery has evolved considerably since Python’s early days. The original specification for packages is still available to read, although some details have changed since the writing of that document.

The original specification for sys.meta_path was PEP 302, with subsequent extension in PEP 420.

PEP 420 introduced namespace packages for Python 3.3. PEP 420 also introduced the find_loader() protocol as an alternative to find_module().

PEP 366 describes the addition of the __package__ attribute for explicit relative imports in main modules.

PEP 328 introduced absolute and explicit relative imports and initially proposed __name__ for semantics PEP 366 would eventually specify for __package__.

PEP 338 defines executing modules as scripts.

PEP 451 adds the encapsulation of per-module import state in spec objects. It also off-loads most of the boilerplate responsibilities of loaders back onto the import machinery. These changes allow the deprecation of several APIs in the import system and also addition of new methods to finders and loaders.

Footnotes

1

See types.ModuleType.

2

The importlib implementation avoids using the return value directly. Instead, it gets the module object by looking the module name up in sys.modules. The indirect effect of this is that an imported module may replace itself in sys.modules. This is implementation-specific behavior that is not guaranteed to work in other Python implementations.

3

In legacy code, it is possible to find instances of imp.NullImporter in the sys.path_importer_cache. It is recommended that code be changed to use None instead. See Porting Python code for more details.

6. Expressions

This chapter explains the meaning of the elements of expressions in Python.

Syntax Notes: In this and the following chapters, extended BNF notation will be used to describe syntax, not lexical analysis. When (one alternative of) a syntax rule has the form

name ::=  othername

and no semantics are given, the semantics of this form of name are the same as for othername.

6.1. Arithmetic conversions

When a description of an arithmetic operator below uses the phrase “the numeric arguments are converted to a common type”, this means that the operator implementation for built-in types works as follows:

  • If either argument is a complex number, the other is converted to complex;

  • otherwise, if either argument is a floating point number, the other is converted to floating point;

  • otherwise, both must be integers and no conversion is necessary.

Some additional rules apply for certain operators (e.g., a string as a left argument to the ‘%’ operator). Extensions must define their own conversion behavior.

6.2. Atoms

Atoms are the most basic elements of expressions. The simplest atoms are identifiers or literals. Forms enclosed in parentheses, brackets or braces are also categorized syntactically as atoms. The syntax for atoms is:

atom      ::=  identifier | literal | enclosure enclosure ::= parenth_form | list_display | dict_display | set_display | generator_expression | yield_atom

6.2.1. Identifiers (Names)

An identifier occurring as an atom is a name. See section Identifiers and keywords for lexical definition and section Naming and binding for documentation of naming and binding.

When the name is bound to an object, evaluation of the atom yields that object. When a name is not bound, an attempt to evaluate it raises a NameError exception.

 

Private name mangling: When an identifier that textually occurs in a class definition begins with two or more underscore characters and does not end in two or more underscores, it is considered a private name of that class. Private names are transformed to a longer form before code is generated for them. The transformation inserts the class name, with leading underscores removed and a single underscore inserted, in front of the name. For example, the identifier __spam occurring in a class named Ham will be transformed to _Ham__spam. This transformation is independent of the syntactical context in which the identifier is used. If the transformed name is extremely long (longer than 255 characters), implementation defined truncation may happen. If the class name consists only of underscores, no transformation is done.

6.2.2. Literals

Python supports string and bytes literals and various numeric literals:

literal ::=  stringliteral | bytesliteral | integer | floatnumber | imagnumber

Evaluation of a literal yields an object of the given type (string, bytes, integer, floating point number, complex number) with the given value. The value may be approximated in the case of floating point and imaginary (complex) literals. See section Literals for details.

All literals correspond to immutable data types, and hence the object’s identity is less important than its value. Multiple evaluations of literals with the same value (either the same occurrence in the program text or a different occurrence) may obtain the same object or a different object with the same value.

6.2.3. Parenthesized forms

A parenthesized form is an optional expression list enclosed in parentheses:

parenth_form ::=  "(" [starred_expression] ")"

A parenthesized expression list yields whatever that expression list yields: if the list contains at least one comma, it yields a tuple; otherwise, it yields the single expression that makes up the expression list.

An empty pair of parentheses yields an empty tuple object. Since tuples are immutable, the same rules as for literals apply (i.e., two occurrences of the empty tuple may or may not yield the same object).

Note that tuples are not formed by the parentheses, but rather by use of the comma operator. The exception is the empty tuple, for which parentheses are required — allowing unparenthesized “nothing” in expressions would cause ambiguities and allow common typos to pass uncaught.

6.2.4. Displays for lists, sets and dictionaries

For constructing a list, a set or a dictionary Python provides special syntax called “displays”, each of them in two flavors:

  • either the container contents are listed explicitly, or

  • they are computed via a set of looping and filtering instructions, called a comprehension.

Common syntax elements for comprehensions are:

comprehension ::=  assignment_expression comp_for comp_for  ::= ["async"] "for" target_list "in" or_test [comp_iter] comp_iter  ::= comp_for | comp_if comp_if  ::= "if" or_test [comp_iter]

The comprehension consists of a single expression followed by at least one for clause and zero or more for or if clauses. In this case, the elements of the new container are those that would be produced by considering each of the for or if clauses a block, nesting from left to right, and evaluating the expression to produce an element each time the innermost block is reached.

However, aside from the iterable expression in the leftmost for clause, the comprehension is executed in a separate implicitly nested scope. This ensures that names assigned to in the target list don’t “leak” into the enclosing scope.

The iterable expression in the leftmost for clause is evaluated directly in the enclosing scope and then passed as an argument to the implicitly nested scope. Subsequent for clauses and any filter condition in the leftmost for clause cannot be evaluated in the enclosing scope as they may depend on the values obtained from the leftmost iterable. For example: [x*y for x in range(10) for y in range(x, x+10)].

To ensure the comprehension always results in a container of the appropriate type, yield and yield from expressions are prohibited in the implicitly nested scope.

Since Python 3.6, in an async def function, an async for clause may be used to iterate over a asynchronous iterator. A comprehension in an async def function may consist of either a for or async for clause following the leading expression, may contain additional for or async for clauses, and may also use await expressions. If a comprehension contains either async for clauses or await expressions it is called an asynchronous comprehension. An asynchronous comprehension may suspend the execution of the coroutine function in which it appears. See also PEP 530.

New in version 3.6: Asynchronous comprehensions were introduced.

Changed in version 3.8: yield and yield from prohibited in the implicitly nested scope.

6.2.5. List displays

A list display is a possibly empty series of expressions enclosed in square brackets:

list_display ::=  "[" [starred_list | comprehension] "]"

A list display yields a new list object, the contents being specified by either a list of expressions or a comprehension. When a comma-separated list of expressions is supplied, its elements are evaluated from left to right and placed into the list object in that order. When a comprehension is supplied, the list is constructed from the elements resulting from the comprehension.

6.2.6. Set displays

A set display is denoted by curly braces and distinguishable from dictionary displays by the lack of colons separating keys and values:

set_display ::=  "{" (starred_list | comprehension) "}"

A set display yields a new mutable set object, the contents being specified by either a sequence of expressions or a comprehension. When a comma-separated list of expressions is supplied, its elements are evaluated from left to right and added to the set object. When a comprehension is supplied, the set is constructed from the elements resulting from the comprehension.

An empty set cannot be constructed with {}; this literal constructs an empty dictionary.

6.2.7. Dictionary displays

A dictionary display is a possibly empty series of key/datum pairs enclosed in curly braces:

dict_display       ::=  "{" [key_datum_list | dict_comprehension] "}" key_datum_list  ::= key_datum ("," key_datum)* [","] key_datum  ::= expression ":" expression | "**" or_expr dict_comprehension ::= expression ":" expression comp_for

A dictionary display yields a new dictionary object.

If a comma-separated sequence of key/datum pairs is given, they are evaluated from left to right to define the entries of the dictionary: each key object is used as a key into the dictionary to store the corresponding datum. This means that you can specify the same key multiple times in the key/datum list, and the final dictionary’s value for that key will be the last one given.

A double asterisk ** denotes dictionary unpacking. Its operand must be a mapping. Each mapping item is added to the new dictionary. Later values replace values already set by earlier key/datum pairs and earlier dictionary unpackings.

New in version 3.5: Unpacking into dictionary displays, originally proposed by PEP 448.

A dict comprehension, in contrast to list and set comprehensions, needs two expressions separated with a colon followed by the usual “for” and “if” clauses. When the comprehension is run, the resulting key and value elements are inserted in the new dictionary in the order they are produced.

Restrictions on the types of the key values are listed earlier in section The standard type hierarchy. (To summarize, the key type should be hashable, which excludes all mutable objects.) Clashes between duplicate keys are not detected; the last datum (textually rightmost in the display) stored for a given key value prevails.

Changed in version 3.8: Prior to Python 3.8, in dict comprehensions, the evaluation order of key and value was not well-defined. In CPython, the value was evaluated before the key. Starting with 3.8, the key is evaluated before the value, as proposed by PEP 572.

6.2.8. Generator expressions

A generator expression is a compact generator notation in parentheses:

generator_expression ::=  "(" expression comp_for ")"

A generator expression yields a new generator object. Its syntax is the same as for comprehensions, except that it is enclosed in parentheses instead of brackets or curly braces.

Variables used in the generator expression are evaluated lazily when the __next__() method is called for the generator object (in the same fashion as normal generators). However, the iterable expression in the leftmost for clause is immediately evaluated, so that an error produced by it will be emitted at the point where the generator expression is defined, rather than at the point where the first value is retrieved. Subsequent for clauses and any filter condition in the leftmost for clause cannot be evaluated in the enclosing scope as they may depend on the values obtained from the leftmost iterable. For example: (x*y for x in range(10) for y in range(x, x+10)).

The parentheses can be omitted on calls with only one argument. See section Calls for details.

To avoid interfering with the expected operation of the generator expression itself, yield and yield from expressions are prohibited in the implicitly defined generator.

If a generator expression contains either async for clauses or await expressions it is called an asynchronous generator expression. An asynchronous generator expression returns a new asynchronous generator object, which is an asynchronous iterator (see Asynchronous Iterators).

New in version 3.6: Asynchronous generator expressions were introduced.

Changed in version 3.7: Prior to Python 3.7, asynchronous generator expressions could only appear in async def coroutines. Starting with 3.7, any function can use asynchronous generator expressions.

Changed in version 3.8: yield and yield from prohibited in the implicitly nested scope.

6.2.9. Yield expressions

yield_atom       ::=  "(" yield_expression ")" yield_expression ::= "yield" [expression_list | "from" expression]

The yield expression is used when defining a generator function or an asynchronous generator function and thus can only be used in the body of a function definition. Using a yield expression in a function’s body causes that function to be a generator function, and using it in an async def function’s body causes that coroutine function to be an asynchronous generator function. For example:

def gen():  # defines a generator function
    yield 123

async def agen(): # defines an asynchronous generator function
    yield 123

Due to their side effects on the containing scope, yield expressions are not permitted as part of the implicitly defined scopes used to implement comprehensions and generator expressions.

Changed in version 3.8: Yield expressions prohibited in the implicitly nested scopes used to implement comprehensions and generator expressions.

Generator functions are described below, while asynchronous generator functions are described separately in section Asynchronous generator functions.

When a generator function is called, it returns an iterator known as a generator. That generator then controls the execution of the generator function. The execution starts when one of the generator’s methods is called. At that time, the execution proceeds to the first yield expression, where it is suspended again, returning the value of expression_list to the generator’s caller. By suspended, we mean that all local state is retained, including the current bindings of local variables, the instruction pointer, the internal evaluation stack, and the state of any exception handling. When the execution is resumed by calling one of the generator’s methods, the function can proceed exactly as if the yield expression were just another external call. The value of the yield expression after resuming depends on the method which resumed the execution. If __next__() is used (typically via either a for or the next() builtin) then the result is None. Otherwise, if send() is used, then the result will be the value passed in to that method.

All of this makes generator functions quite similar to coroutines; they yield multiple times, they have more than one entry point and their execution can be suspended. The only difference is that a generator function cannot control where the execution should continue after it yields; the control is always transferred to the generator’s caller.

Yield expressions are allowed anywhere in a try construct. If the generator is not resumed before it is finalized (by reaching a zero reference count or by being garbage collected), the generator-iterator’s close() method will be called, allowing any pending finally clauses to execute.

When yield from  is used, the supplied expression must be an iterable. The values produced by iterating that iterable are passed directly to the caller of the current generator’s methods. Any values passed in with send() and any exceptions passed in with throw() are passed to the underlying iterator if it has the appropriate methods. If this is not the case, then send() will raise AttributeError or TypeError, while throw() will just raise the passed in exception immediately.

When the underlying iterator is complete, the value attribute of the raised StopIteration instance becomes the value of the yield expression. It can be either set explicitly when raising StopIteration, or automatically when the subiterator is a generator (by returning a value from the subgenerator).

Changed in version 3.3: Added yield from  to delegate control flow to a subiterator.

The parentheses may be omitted when the yield expression is the sole expression on the right hand side of an assignment statement.

See also

PEP 255 – Simple Generators

The proposal for adding generators and the yield statement to Python.

PEP 342 – Coroutines via Enhanced Generators

The proposal to enhance the API and syntax of generators, making them usable as simple coroutines.

PEP 380 – Syntax for Delegating to a Subgenerator

The proposal to introduce the yield_from syntax, making delegation to subgenerators easy.

PEP 525 – Asynchronous Generators

The proposal that expanded on PEP 492 by adding generator capabilities to coroutine functions.

6.2.9.1. Generator-iterator methods

This subsection describes the methods of a generator iterator. They can be used to control the execution of a generator function.

Note that calling any of the generator methods below when the generator is already executing raises a ValueError exception.

 

generator.__next__()

Starts the execution of a generator function or resumes it at the last executed yield expression. When a generator function is resumed with a __next__() method, the current yield expression always evaluates to None. The execution then continues to the next yield expression, where the generator is suspended again, and the value of the expression_list is returned to __next__()’s caller. If the generator exits without yielding another value, a StopIteration exception is raised.

This method is normally called implicitly, e.g. by a for loop, or by the built-in next() function.

generator.send(value)

Resumes the execution and “sends” a value into the generator function. The value argument becomes the result of the current yield expression. The send() method returns the next value yielded by the generator, or raises StopIteration if the generator exits without yielding another value. When send() is called to start the generator, it must be called with None as the argument, because there is no yield expression that could receive the value.

generator.throw(value)
generator.throw(type[value[traceback]])

Raises an exception at the point where the generator was paused, and returns the next value yielded by the generator function. If the generator exits without yielding another value, a StopIteration exception is raised. If the generator function does not catch the passed-in exception, or raises a different exception, then that exception propagates to the caller.

In typical use, this is called with a single exception instance similar to the way the raise keyword is used.

For backwards compatability, however, the second signature is supported, following a convention from older versions of Python. The type argument should be an exception class, and value should be an exception instance. If the value is not provided, the type constructor is called to get an instance. If traceback is provided, it is set on the exception, otherwise any existing __traceback__ attribute stored in value may be cleared.

 

generator.close()

Raises a GeneratorExit at the point where the generator function was paused. If the generator function then exits gracefully, is already closed, or raises GeneratorExit (by not catching the exception), close returns to its caller. If the generator yields a value, a RuntimeError is raised. If the generator raises any other exception, it is propagated to the caller. close() does nothing if the generator has already exited due to an exception or normal exit.

6.2.9.2. Examples

Here is a simple example that demonstrates the behavior of generators and generator functions:

>>>

>>> def echo(value=None):
...     print("Execution starts when 'next()' is called for the first time.")
...     try:
...         while True:
...             try:
...                 value = (yield value)
...             except Exception as e:
...                 value = e
...     finally:
...         print("Don't forget to clean up when 'close()' is called.")
...
>>> generator = echo(1)
>>> print(next(generator))
Execution starts when 'next()' is called for the first time.
1
>>> print(next(generator))
None
>>> print(generator.send(2))
2
>>> generator.throw(TypeError, "spam")
TypeError('spam',)
>>> generator.close()
Don't forget to clean up when 'close()' is called.

For examples using yield from, see PEP 380: Syntax for Delegating to a Subgenerator in “What’s New in Python.”

6.2.9.3. Asynchronous generator functions

The presence of a yield expression in a function or method defined using async def further defines the function as an asynchronous generator function.

When an asynchronous generator function is called, it returns an asynchronous iterator known as an asynchronous generator object. That object then controls the execution of the generator function. An asynchronous generator object is typically used in an async for statement in a coroutine function analogously to how a generator object would be used in a for statement.

Calling one of the asynchronous generator’s methods returns an awaitable object, and the execution starts when this object is awaited on. At that time, the execution proceeds to the first yield expression, where it is suspended again, returning the value of expression_list to the awaiting coroutine. As with a generator, suspension means that all local state is retained, including the current bindings of local variables, the instruction pointer, the internal evaluation stack, and the state of any exception handling. When the execution is resumed by awaiting on the next object returned by the asynchronous generator’s methods, the function can proceed exactly as if the yield expression were just another external call. The value of the yield expression after resuming depends on the method which resumed the execution. If __anext__() is used then the result is None. Otherwise, if asend() is used, then the result will be the value passed in to that method.

If an asynchronous generator happens to exit early by break, the caller task being cancelled, or other exceptions, the generator’s async cleanup code will run and possibly raise exceptions or access context variables in an unexpected context–perhaps after the lifetime of tasks it depends, or during the event loop shutdown when the async-generator garbage collection hook is called. To prevent this, the caller must explicitly close the async generator by calling aclose() method to finalize the generator and ultimately detach it from the event loop.

In an asynchronous generator function, yield expressions are allowed anywhere in a try construct. However, if an asynchronous generator is not resumed before it is finalized (by reaching a zero reference count or by being garbage collected), then a yield expression within a try construct could result in a failure to execute pending finally clauses. In this case, it is the responsibility of the event loop or scheduler running the asynchronous generator to call the asynchronous generator-iterator’s aclose() method and run the resulting coroutine object, thus allowing any pending finally clauses to execute.

To take care of finalization upon event loop termination, an event loop should define a finalizer function which takes an asynchronous generator-iterator and presumably calls aclose() and executes the coroutine. This finalizer may be registered by calling sys.set_asyncgen_hooks(). When first iterated over, an asynchronous generator-iterator will store the registered finalizer to be called upon finalization. For a reference example of a finalizer method see the implementation of asyncio.Loop.shutdown_asyncgens in Lib/asyncio/base_events.py.

The expression yield from  is a syntax error when used in an asynchronous generator function.

6.2.9.4. Asynchronous generator-iterator methods

This subsection describes the methods of an asynchronous generator iterator, which are used to control the execution of a generator function.

 

coroutine agen.__anext__()

Returns an awaitable which when run starts to execute the asynchronous generator or resumes it at the last executed yield expression. When an asynchronous generator function is resumed with an __anext__() method, the current yield expression always evaluates to None in the returned awaitable, which when run will continue to the next yield expression. The value of the expression_list of the yield expression is the value of the StopIteration exception raised by the completing coroutine. If the asynchronous generator exits without yielding another value, the awaitable instead raises a StopAsyncIteration exception, signalling that the asynchronous iteration has completed.

This method is normally called implicitly by a async for loop.

coroutine agen.asend(value)

Returns an awaitable which when run resumes the execution of the asynchronous generator. As with the send() method for a generator, this “sends” a value into the asynchronous generator function, and the value argument becomes the result of the current yield expression. The awaitable returned by the asend() method will return the next value yielded by the generator as the value of the raised StopIteration, or raises StopAsyncIteration if the asynchronous generator exits without yielding another value. When asend() is called to start the asynchronous generator, it must be called with None as the argument, because there is no yield expression that could receive the value.

coroutine agen.athrow(type[value[traceback]])

Returns an awaitable that raises an exception of type type at the point where the asynchronous generator was paused, and returns the next value yielded by the generator function as the value of the raised StopIteration exception. If the asynchronous generator exits without yielding another value, a StopAsyncIteration exception is raised by the awaitable. If the generator function does not catch the passed-in exception, or raises a different exception, then when the awaitable is run that exception propagates to the caller of the awaitable.

 

coroutine agen.aclose()

Returns an awaitable that when run will throw a GeneratorExit into the asynchronous generator function at the point where it was paused. If the asynchronous generator function then exits gracefully, is already closed, or raises GeneratorExit (by not catching the exception), then the returned awaitable will raise a StopIteration exception. Any further awaitables returned by subsequent calls to the asynchronous generator will raise a StopAsyncIteration exception. If the asynchronous generator yields a value, a RuntimeError is raised by the awaitable. If the asynchronous generator raises any other exception, it is propagated to the caller of the awaitable. If the asynchronous generator has already exited due to an exception or normal exit, then further calls to aclose() will return an awaitable that does nothing.

6.3. Primaries

Primaries represent the most tightly bound operations of the language. Their syntax is:

primary ::=  atom | attributeref | subscription | slicing | call

6.3.1. Attribute references

An attribute reference is a primary followed by a period and a name:

attributeref ::=  primary "." identifier

The primary must evaluate to an object of a type that supports attribute references, which most objects do. This object is then asked to produce the attribute whose name is the identifier. This production can be customized by overriding the __getattr__() method. If this attribute is not available, the exception AttributeError is raised. Otherwise, the type and value of the object produced is determined by the object. Multiple evaluations of the same attribute reference may yield different objects.

6.3.2. Subscriptions

 

The subscription of an instance of a container class will generally select an element from the container. The subscription of a generic class will generally return a GenericAlias object.

subscription ::=  primary "[" expression_list "]"

When an object is subscripted, the interpreter will evaluate the primary and the expression list.

The primary must evaluate to an object that supports subscription. An object may support subscription through defining one or both of __getitem__() and __class_getitem__(). When the primary is subscripted, the evaluated result of the expression list will be passed to one of these methods. For more details on when __class_getitem__ is called instead of __getitem__, see __class_getitem__ versus __getitem__.

If the expression list contains at least one comma, it will evaluate to a tuple containing the items of the expression list. Otherwise, the expression list will evaluate to the value of the list’s sole member.

For built-in objects, there are two types of objects that support subscription via __getitem__():

  1. Mappings. If the primary is a mapping, the expression list must evaluate to an object whose value is one of the keys of the mapping, and the subscription selects the value in the mapping that corresponds to that key. An example of a builtin mapping class is the dict class.

  2. Sequences. If the primary is a sequence, the expression list must evaluate to an int or a slice (as discussed in the following section). Examples of builtin sequence classes include the strlist and tuple classes.

The formal syntax makes no special provision for negative indices in sequences. However, built-in sequences all provide a __getitem__() method that interprets negative indices by adding the length of the sequence to the index so that, for example, x[-1] selects the last item of x. The resulting value must be a nonnegative integer less than the number of items in the sequence, and the subscription selects the item whose index is that value (counting from zero). Since the support for negative indices and slicing occurs in the object’s __getitem__() method, subclasses overriding this method will need to explicitly add that support.

string is a special kind of sequence whose items are characters. A character is not a separate data type but a string of exactly one character.

6.3.3. Slicings

 

A slicing selects a range of items in a sequence object (e.g., a string, tuple or list). Slicings may be used as expressions or as targets in assignment or del statements. The syntax for a slicing:

slicing      ::=  primary "[" slice_list "]" slice_list  ::= slice_item ("," slice_item)* [","] slice_item  ::= expression | proper_slice proper_slice ::= [lower_bound] ":" [upper_bound] [ ":" [stride] ] lower_bound  ::= expression upper_bound  ::= expression stride  ::= expression

There is ambiguity in the formal syntax here: anything that looks like an expression list also looks like a slice list, so any subscription can be interpreted as a slicing. Rather than further complicating the syntax, this is disambiguated by defining that in this case the interpretation as a subscription takes priority over the interpretation as a slicing (this is the case if the slice list contains no proper slice).

The semantics for a slicing are as follows. The primary is indexed (using the same __getitem__() method as normal subscription) with a key that is constructed from the slice list, as follows. If the slice list contains at least one comma, the key is a tuple containing the conversion of the slice items; otherwise, the conversion of the lone slice item is the key. The conversion of a slice item that is an expression is that expression. The conversion of a proper slice is a slice object (see section The standard type hierarchy) whose startstop and step attributes are the values of the expressions given as lower bound, upper bound and stride, respectively, substituting None for missing expressions.

6.3.4. Calls

A call calls a callable object (e.g., a function) with a possibly empty series of arguments:

call                 ::=  primary "(" [argument_list [","] | comprehension] ")" argument_list  ::= positional_arguments ["," starred_and_keywords] ["," keywords_arguments] | starred_and_keywords ["," keywords_arguments] | keywords_arguments positional_arguments ::= positional_item ("," positional_item)* positional_item  ::= assignment_expression | "*" expression starred_and_keywords ::= ("*" expression | keyword_item) ("," "*" expression | "," keyword_item)* keywords_arguments  ::= (keyword_item | "**" expression) ("," keyword_item | "," "**" expression)* keyword_item  ::= identifier "=" expression

An optional trailing comma may be present after the positional and keyword arguments but does not affect the semantics.

The primary must evaluate to a callable object (user-defined functions, built-in functions, methods of built-in objects, class objects, methods of class instances, and all objects having a __call__() method are callable). All argument expressions are evaluated before the call is attempted. Please refer to section Function definitions for the syntax of formal parameter lists.

If keyword arguments are present, they are first converted to positional arguments, as follows. First, a list of unfilled slots is created for the formal parameters. If there are N positional arguments, they are placed in the first N slots. Next, for each keyword argument, the identifier is used to determine the corresponding slot (if the identifier is the same as the first formal parameter name, the first slot is used, and so on). If the slot is already filled, a TypeError exception is raised. Otherwise, the value of the argument is placed in the slot, filling it (even if the expression is None, it fills the slot). When all arguments have been processed, the slots that are still unfilled are filled with the corresponding default value from the function definition. (Default values are calculated, once, when the function is defined; thus, a mutable object such as a list or dictionary used as default value will be shared by all calls that don’t specify an argument value for the corresponding slot; this should usually be avoided.) If there are any unfilled slots for which no default value is specified, a TypeError exception is raised. Otherwise, the list of filled slots is used as the argument list for the call.

CPython implementation detail: An implementation may provide built-in functions whose positional parameters do not have names, even if they are ‘named’ for the purpose of documentation, and which therefore cannot be supplied by keyword. In CPython, this is the case for functions implemented in C that use PyArg_ParseTuple() to parse their arguments.

If there are more positional arguments than there are formal parameter slots, a TypeError exception is raised, unless a formal parameter using the syntax *identifier is present; in this case, that formal parameter receives a tuple containing the excess positional arguments (or an empty tuple if there were no excess positional arguments).

If any keyword argument does not correspond to a formal parameter name, a TypeError exception is raised, unless a formal parameter using the syntax **identifier is present; in this case, that formal parameter receives a dictionary containing the excess keyword arguments (using the keywords as keys and the argument values as corresponding values), or a (new) empty dictionary if there were no excess keyword arguments.

If the syntax *expression appears in the function call, expression must evaluate to an iterable. Elements from these iterables are treated as if they were additional positional arguments. For the call f(x1, x2, *y, x3, x4), if y evaluates to a sequence y1, …, yM, this is equivalent to a call with M+4 positional arguments x1x2y1, …, yMx3x4.

A consequence of this is that although the *expression syntax may appear after explicit keyword arguments, it is processed before the keyword arguments (and any **expression arguments – see below). So:

>>>

>>> def f(a, b):
...     print(a, b)
...
>>> f(b=1, *(2,))
2 1
>>> f(a=1, *(2,))
Traceback (most recent call last):
  File "", line 1, in 
TypeError: f() got multiple values for keyword argument 'a'
>>> f(1, *(2,))
1 2

It is unusual for both keyword arguments and the *expression syntax to be used in the same call, so in practice this confusion does not arise.

If the syntax **expression appears in the function call, expression must evaluate to a mapping, the contents of which are treated as additional keyword arguments. If a keyword is already present (as an explicit keyword argument, or from another unpacking), a TypeError exception is raised.

Formal parameters using the syntax *identifier or **identifier cannot be used as positional argument slots or as keyword argument names.

Changed in version 3.5: Function calls accept any number of * and ** unpackings, positional arguments may follow iterable unpackings (*), and keyword arguments may follow dictionary unpackings (**). Originally proposed by PEP 448.

A call always returns some value, possibly None, unless it raises an exception. How this value is computed depends on the type of the callable object.

If it is—

a user-defined function:

The code block for the function is executed, passing it the argument list. The first thing the code block will do is bind the formal parameters to the arguments; this is described in section Function definitions. When the code block executes a return statement, this specifies the return value of the function call.

a built-in function or method:

The result is up to the interpreter; see Built-in Functions for the descriptions of built-in functions and methods.

a class object:

A new instance of that class is returned.

a class instance method:

The corresponding user-defined function is called, with an argument list that is one longer than the argument list of the call: the instance becomes the first argument.

a class instance:

The class must define a __call__() method; the effect is then the same as if that method was called.

6.4. Await expression

Suspend the execution of coroutine on an awaitable object. Can only be used inside a coroutine function.

await_expr ::=  "await" primary

New in version 3.5.

6.5. The power operator

The power operator binds more tightly than unary operators on its left; it binds less tightly than unary operators on its right. The syntax is:

power ::=  (await_expr | primary) ["**" u_expr]

Thus, in an unparenthesized sequence of power and unary operators, the operators are evaluated from right to left (this does not constrain the evaluation order for the operands): -1**2 results in -1.

The power operator has the same semantics as the built-in pow() function, when called with two arguments: it yields its left argument raised to the power of its right argument. The numeric arguments are first converted to a common type, and the result is of that type.

For int operands, the result has the same type as the operands unless the second argument is negative; in that case, all arguments are converted to float and a float result is delivered. For example, 10**2 returns 100, but 10**-2 returns 0.01.

Raising 0.0 to a negative power results in a ZeroDivisionError. Raising a negative number to a fractional power results in a complex number. (In earlier versions it raised a ValueError.)

This operation can be customized using the special __pow__() method.

6.6. Unary arithmetic and bitwise operations

All unary arithmetic and bitwise operations have the same priority:

u_expr ::=  power | "-" u_expr | "+" u_expr | "~" u_expr

The unary - (minus) operator yields the negation of its numeric argument; the operation can be overridden with the __neg__() special method.

The unary + (plus) operator yields its numeric argument unchanged; the operation can be overridden with the __pos__() special method.

The unary ~ (invert) operator yields the bitwise inversion of its integer argument. The bitwise inversion of x is defined as -(x+1). It only applies to integral numbers or to custom objects that override the __invert__() special method.

In all three cases, if the argument does not have the proper type, a TypeError exception is raised.

6.7. Binary arithmetic operations

The binary arithmetic operations have the conventional priority levels. Note that some of these operations also apply to certain non-numeric types. Apart from the power operator, there are only two levels, one for multiplicative operators and one for additive operators:

m_expr ::=  u_expr | m_expr "*" u_expr | m_expr "@" m_expr | m_expr "//" u_expr | m_expr "/" u_expr | m_expr "%" u_expr a_expr ::= m_expr | a_expr "+" m_expr | a_expr "-" m_expr

The * (multiplication) operator yields the product of its arguments. The arguments must either both be numbers, or one argument must be an integer and the other must be a sequence. In the former case, the numbers are converted to a common type and then multiplied together. In the latter case, sequence repetition is performed; a negative repetition factor yields an empty sequence.

This operation can be customized using the special __mul__() and __rmul__() methods.

The @ (at) operator is intended to be used for matrix multiplication. No builtin Python types implement this operator.

New in version 3.5.

The / (division) and // (floor division) operators yield the quotient of their arguments. The numeric arguments are first converted to a common type. Division of integers yields a float, while floor division of integers results in an integer; the result is that of mathematical division with the ‘floor’ function applied to the result. Division by zero raises the ZeroDivisionError exception.

This operation can be customized using the special __truediv__() and __floordiv__() methods.

The % (modulo) operator yields the remainder from the division of the first argument by the second. The numeric arguments are first converted to a common type. A zero right argument raises the ZeroDivisionError exception. The arguments may be floating point numbers, e.g., 3.14%0.7 equals 0.34 (since 3.14 equals 4*0.7 + 0.34.) The modulo operator always yields a result with the same sign as its second operand (or zero); the absolute value of the result is strictly smaller than the absolute value of the second operand 1.

The floor division and modulo operators are connected by the following identity: x == (x//y)*y + (x%y). Floor division and modulo are also connected with the built-in function divmod()divmod(x, y) == (x//y, x%y)2.

In addition to performing the modulo operation on numbers, the % operator is also overloaded by string objects to perform old-style string formatting (also known as interpolation). The syntax for string formatting is described in the Python Library Reference, section printf-style String Formatting.

The modulo operation can be customized using the special __mod__() method.

The floor division operator, the modulo operator, and the divmod() function are not defined for complex numbers. Instead, convert to a floating point number using the abs() function if appropriate.

The + (addition) operator yields the sum of its arguments. The arguments must either both be numbers or both be sequences of the same type. In the former case, the numbers are converted to a common type and then added together. In the latter case, the sequences are concatenated.

This operation can be customized using the special __add__() and __radd__() methods.

The - (subtraction) operator yields the difference of its arguments. The numeric arguments are first converted to a common type.

This operation can be customized using the special __sub__() method.

6.8. Shifting operations

The shifting operations have lower priority than the arithmetic operations:

shift_expr ::=  a_expr | shift_expr ("<<" | ">>") a_expr

These operators accept integers as arguments. They shift the first argument to the left or right by the number of bits given by the second argument.

This operation can be customized using the special __lshift__() and __rshift__() methods.

A right shift by n bits is defined as floor division by pow(2,n). A left shift by n bits is defined as multiplication with pow(2,n).

6.9. Binary bitwise operations

Each of the three bitwise operations has a different priority level:

and_expr ::=  shift_expr | and_expr "&" shift_expr xor_expr ::= and_expr | xor_expr "^" and_expr or_expr  ::= xor_expr | or_expr "|" xor_expr

The & operator yields the bitwise AND of its arguments, which must be integers or one of them must be a custom object overriding __and__() or __rand__() special methods.

The ^ operator yields the bitwise XOR (exclusive OR) of its arguments, which must be integers or one of them must be a custom object overriding __xor__() or __rxor__() special methods.

The | operator yields the bitwise (inclusive) OR of its arguments, which must be integers or one of them must be a custom object overriding __or__() or __ror__() special methods.

6.10. Comparisons

Unlike C, all comparison operations in Python have the same priority, which is lower than that of any arithmetic, shifting or bitwise operation. Also unlike C, expressions like a < b < c have the interpretation that is conventional in mathematics:

comparison    ::=  or_expr (comp_operator or_expr)* comp_operator ::= "<" | ">" | "==" | ">=" | "<=" | "!=" | "is" ["not"] | ["not"] "in"

Comparisons yield boolean values: True or False. Custom rich comparison methods may return non-boolean values. In this case Python will call bool() on such value in boolean contexts.

Comparisons can be chained arbitrarily, e.g., x < y <= z is equivalent to x < y and y <= z, except that y is evaluated only once (but in both cases z is not evaluated at all when x < y is found to be false).

Formally, if abc, …, yz are expressions and op1op2, …, opN are comparison operators, then a op1 b op2 c ... y opN z is equivalent to a op1 b and b op2 c and ... y opN z, except that each expression is evaluated at most once.

Note that a op1 b op2 c doesn’t imply any kind of comparison between a and c, so that, e.g., x < y > z is perfectly legal (though perhaps not pretty).

6.10.1. Value comparisons

The operators <>==>=<=, and != compare the values of two objects. The objects do not need to have the same type.

Chapter Objects, values and types states that objects have a value (in addition to type and identity). The value of an object is a rather abstract notion in Python: For example, there is no canonical access method for an object’s value. Also, there is no requirement that the value of an object should be constructed in a particular way, e.g. comprised of all its data attributes. Comparison operators implement a particular notion of what the value of an object is. One can think of them as defining the value of an object indirectly, by means of their comparison implementation.

Because all types are (direct or indirect) subtypes of object, they inherit the default comparison behavior from object. Types can customize their comparison behavior by implementing rich comparison methods like __lt__(), described in Basic customization.

The default behavior for equality comparison (== and !=) is based on the identity of the objects. Hence, equality comparison of instances with the same identity results in equality, and equality comparison of instances with different identities results in inequality. A motivation for this default behavior is the desire that all objects should be reflexive (i.e. x is y implies x == y).

A default order comparison (<><=, and >=) is not provided; an attempt raises TypeError. A motivation for this default behavior is the lack of a similar invariant as for equality.

The behavior of the default equality comparison, that instances with different identities are always unequal, may be in contrast to what types will need that have a sensible definition of object value and value-based equality. Such types will need to customize their comparison behavior, and in fact, a number of built-in types have done that.

The following list describes the comparison behavior of the most important built-in types.

  • Numbers of built-in numeric types (Numeric Types — int, float, complex) and of the standard library types fractions.Fraction and decimal.Decimal can be compared within and across their types, with the restriction that complex numbers do not support order comparison. Within the limits of the types involved, they compare mathematically (algorithmically) correct without loss of precision.

    The not-a-number values float('NaN') and decimal.Decimal('NaN') are special. Any ordered comparison of a number to a not-a-number value is false. A counter-intuitive implication is that not-a-number values are not equal to themselves. For example, if x = float('NaN')3 < xx < 3 and x == x are all false, while x != x is true. This behavior is compliant with IEEE 754.

  • None and NotImplemented are singletons. PEP 8 advises that comparisons for singletons should always be done with is or is not, never the equality operators.

  • Binary sequences (instances of bytes or bytearray) can be compared within and across their types. They compare lexicographically using the numeric values of their elements.

  • Strings (instances of str) compare lexicographically using the numerical Unicode code points (the result of the built-in function ord()) of their characters. 3

    Strings and binary sequences cannot be directly compared.

  • Sequences (instances of tuplelist, or range) can be compared only within each of their types, with the restriction that ranges do not support order comparison. Equality comparison across these types results in inequality, and ordering comparison across these types raises TypeError.

    Sequences compare lexicographically using comparison of corresponding elements. The built-in containers typically assume identical objects are equal to themselves. That lets them bypass equality tests for identical objects to improve performance and to maintain their internal invariants.

    Lexicographical comparison between built-in collections works as follows:

    • For two collections to compare equal, they must be of the same type, have the same length, and each pair of corresponding elements must compare equal (for example, [1,2] == (1,2) is false because the type is not the same).

    • Collections that support order comparison are ordered the same as their first unequal elements (for example, [1,2,x] <= [1,2,y] has the same value as x <= y). If a corresponding element does not exist, the shorter collection is ordered first (for example, [1,2] < [1,2,3] is true).

  • Mappings (instances of dict) compare equal if and only if they have equal (key, value) pairs. Equality comparison of the keys and values enforces reflexivity.

    Order comparisons (<><=, and >=) raise TypeError.

  • Sets (instances of set or frozenset) can be compared within and across their types.

    They define order comparison operators to mean subset and superset tests. Those relations do not define total orderings (for example, the two sets {1,2} and {2,3} are not equal, nor subsets of one another, nor supersets of one another). Accordingly, sets are not appropriate arguments for functions which depend on total ordering (for example, min()max(), and sorted() produce undefined results given a list of sets as inputs).

    Comparison of sets enforces reflexivity of its elements.

  • Most other built-in types have no comparison methods implemented, so they inherit the default comparison behavior.

User-defined classes that customize their comparison behavior should follow some consistency rules, if possible:

  • Equality comparison should be reflexive. In other words, identical objects should compare equal:

    x is y implies x == y

  • Comparison should be symmetric. In other words, the following expressions should have the same result:

    x == y and y == x

    x != y and y != x

    x < y and y > x

    x <= y and y >= x

  • Comparison should be transitive. The following (non-exhaustive) examples illustrate that:

    x > y and y > z implies x > z

    x < y and y <= z implies x < z

  • Inverse comparison should result in the boolean negation. In other words, the following expressions should have the same result:

    x == y and not x != y

    x < y and not x >= y (for total ordering)

    x > y and not x <= y (for total ordering)

    The last two expressions apply to totally ordered collections (e.g. to sequences, but not to sets or mappings). See also the total_ordering() decorator.

  • The hash() result should be consistent with equality. Objects that are equal should either have the same hash value, or be marked as unhashable.

Python does not enforce these consistency rules. In fact, the not-a-number values are an example for not following these rules.

6.10.2. Membership test operations

The operators in and not in test for membership. x in s evaluates to True if x is a member of s, and False otherwise. x not in s returns the negation of x in s. All built-in sequences and set types support this as well as dictionary, for which in tests whether the dictionary has a given key. For container types such as list, tuple, set, frozenset, dict, or collections.deque, the expression x in y is equivalent to any(x is e or x == e for e in y).

For the string and bytes types, x in y is True if and only if x is a substring of y. An equivalent test is y.find(x) != -1. Empty strings are always considered to be a substring of any other string, so "" in "abc" will return True.

For user-defined classes which define the __contains__() method, x in y returns True if y.__contains__(x) returns a true value, and False otherwise.

For user-defined classes which do not define __contains__() but do define __iter__()x in y is True if some value z, for which the expression x is z or x == z is true, is produced while iterating over y. If an exception is raised during the iteration, it is as if in raised that exception.

Lastly, the old-style iteration protocol is tried: if a class defines __getitem__()x in y is True if and only if there is a non-negative integer index i such that x is y[i] or x == y[i], and no lower integer index raises the IndexError exception. (If any other exception is raised, it is as if in raised that exception).

The operator not in is defined to have the inverse truth value of in.

6.10.3. Identity comparisons

The operators is and is not test for an object’s identity: x is y is true if and only if x and y are the same object. An Object’s identity is determined using the id() function. x is not y yields the inverse truth value. 4

6.11. Boolean operations

or_test  ::=  and_test | or_test "or" and_test and_test ::= not_test | and_test "and" not_test not_test ::= comparison | "not" not_test

In the context of Boolean operations, and also when expressions are used by control flow statements, the following values are interpreted as false: FalseNone, numeric zero of all types, and empty strings and containers (including strings, tuples, lists, dictionaries, sets and frozensets). All other values are interpreted as true. User-defined objects can customize their truth value by providing a __bool__() method.

The operator not yields True if its argument is false, False otherwise.

The expression x and y first evaluates x; if x is false, its value is returned; otherwise, y is evaluated and the resulting value is returned.

The expression x or y first evaluates x; if x is true, its value is returned; otherwise, y is evaluated and the resulting value is returned.

Note that neither and nor or restrict the value and type they return to False and True, but rather return the last evaluated argument. This is sometimes useful, e.g., if s is a string that should be replaced by a default value if it is empty, the expression s or 'foo' yields the desired value. Because not has to create a new value, it returns a boolean value regardless of the type of its argument (for example, not 'foo' produces False rather than ''.)

6.12. Assignment expressions

assignment_expression ::=  [identifier ":="] expression

An assignment expression (sometimes also called a “named expression” or “walrus”) assigns an expression to an identifier, while also returning the value of the expression.

One common use case is when handling matched regular expressions:

if matching := pattern.search(data):
    do_something(matching)

Or, when processing a file stream in chunks:

while chunk := file.read(9000):
    process(chunk)

New in version 3.8: See PEP 572 for more details about assignment expressions.

6.13. Conditional expressions

conditional_expression ::=  or_test ["if" or_test "else" expression] expression  ::= conditional_expression | lambda_expr

Conditional expressions (sometimes called a “ternary operator”) have the lowest priority of all Python operations.

The expression x if C else y first evaluates the condition, C rather than x. If C is true, x is evaluated and its value is returned; otherwise, y is evaluated and its value is returned.

See PEP 308 for more details about conditional expressions.

6.14. Lambdas

lambda_expr ::=  "lambda" [parameter_list] ":" expression

Lambda expressions (sometimes called lambda forms) are used to create anonymous functions. The expression lambda parameters: expression yields a function object. The unnamed object behaves like a function object defined with:

def (parameters):
    return expression

See section Function definitions for the syntax of parameter lists. Note that functions created with lambda expressions cannot contain statements or annotations.

6.15. Expression lists

expression_list    ::=  expression ("," expression)* [","] starred_list  ::= starred_item ("," starred_item)* [","] starred_expression ::= expression | (starred_item ",")* [starred_item] starred_item  ::= assignment_expression | "*" or_expr

Except when part o