Update

This PEP has partially been superseded by PEP 3137.

Abstract

This PEP outlines the introduction of a raw bytes sequence type. Adding the bytes type is one step in the transition to Unicode-based str objects which will be introduced in Python 3.0.

The PEP describes how the bytes type should work in Python 2.6, as well as how it should work in Python 3.0. (Occasionally there are differences because in Python 2.6, we have two string types, str and unicode, while in Python 3.0 we will only have one string type, whose name will be str but whose semantics will be like the 2.6 unicode type.)

Motivation

Python’s current string objects are overloaded. They serve to hold both sequences of characters and sequences of bytes. This overloading of purpose leads to confusion and bugs. In future versions of Python, string objects will be used for holding character data. The bytes object will fulfil the role of a byte container. Eventually the unicode type will be renamed to str and the old str type will be removed.

Specification

A bytes object stores a mutable sequence of integers that are in the range 0 to 255. Unlike string objects, indexing a bytes object returns an integer. Assigning or comparing an object that is not an integer to an element causes a TypeError exception. Assigning an element to a value outside the range 0 to 255 causes a ValueError exception. The .__len__() method of bytes returns the number of integers stored in the sequence (i.e. the number of bytes).

The constructor of the bytes object has the following signature:

bytes([initializer[, encoding]])

If no arguments are provided then a bytes object containing zero elements is created and returned. The initializer argument can be a string (in 2.6, either str or unicode), an iterable of integers, or a single integer. The pseudo-code for the constructor (optimized for clear semantics, not for speed) is:

def bytes(initializer=0, encoding=None):
    if isinstance(initializer, int): # In 2.6, int -> (int, long)
        initializer = [0]*initializer
    elif isinstance(initializer, basestring):
        if isinstance(initializer, unicode): # In 3.0, "if True"
            if encoding is None:
                # In 3.0, raise TypeError("explicit encoding required")
                encoding = sys.getdefaultencoding()
            initializer = initializer.encode(encoding)
        initializer = [ord(c) for c in initializer]
    else:
        if encoding is not None:
            raise TypeError("no encoding allowed for this initializer")
        tmp = []
        for c in initializer:
            if not isinstance(c, int):
                raise TypeError("initializer must be iterable of ints")
            if not 0 <= c < 256:
                raise ValueError("initializer element out of range")
            tmp.append(c)
        initializer = tmp
    new = <new bytes object of length len(initializer)>
    for i, c in enumerate(initializer):
        new[i] = c
    return new

The .__repr__() method returns a string that can be evaluated to generate a new bytes object containing a bytes literal:

>>> bytes([10, 20, 30])
b'\n\x14\x1e'

The object has a .decode() method equivalent to the .decode() method of the str object. The object has a classmethod .fromhex() that takes a string of characters from the set [0-9a-fA-F ] and returns a bytes object (similar to binascii.unhexlify). For example:

>>> bytes.fromhex('5c5350ff')
b'\\SP\xff'
>>> bytes.fromhex('5c 53 50 ff')
b'\\SP\xff'

The object has a .hex() method that does the reverse conversion (similar to binascii.hexlify):

>> bytes([92, 83, 80, 255]).hex()
'5c5350ff'

The bytes object has some methods similar to list methods, and others similar to str methods. Here is a complete list of methods, with their approximate signatures:

.__add__(bytes) -> bytes
.__contains__(int | bytes) -> bool
.__delitem__(int | slice) -> None
.__delslice__(int, int) -> None
.__eq__(bytes) -> bool
.__ge__(bytes) -> bool
.__getitem__(int | slice) -> int | bytes
.__getslice__(int, int) -> bytes
.__gt__(bytes) -> bool
.__iadd__(bytes) -> bytes
.__imul__(int) -> bytes
.__iter__() -> iterator
.__le__(bytes) -> bool
.__len__() -> int
.__lt__(bytes) -> bool
.__mul__(int) -> bytes
.__ne__(bytes) -> bool
.__reduce__(...) -> ...
.__reduce_ex__(...) -> ...
.__repr__() -> str
.__reversed__() -> bytes
.__rmul__(int) -> bytes
.__setitem__(int | slice, int | iterable[int]) -> None
.__setslice__(int, int, iterable[int]) -> Bote
.append(int) -> None
.count(int) -> int
.decode(str) -> str | unicode # in 3.0, only str
.endswith(bytes) -> bool
.extend(iterable[int]) -> None
.find(bytes) -> int
.index(bytes | int) -> int
.insert(int, int) -> None
.join(iterable[bytes]) -> bytes
.partition(bytes) -> (bytes, bytes, bytes)
.pop([int]) -> int
.remove(int) -> None
.replace(bytes, bytes) -> bytes
.rindex(bytes | int) -> int
.rpartition(bytes) -> (bytes, bytes, bytes)
.split(bytes) -> list[bytes]
.startswith(bytes) -> bool
.reverse() -> None
.rfind(bytes) -> int
.rindex(bytes | int) -> int
.rsplit(bytes) -> list[bytes]
.translate(bytes, [bytes]) -> bytes

Note the conspicuous absence of .isupper(), .upper(), and friends. (But see “Open Issues” below.) There is no .__hash__() because the object is mutable. There is no use case for a .sort() method.

The bytes type also supports the buffer interface, supporting reading and writing binary (but not character) data.

Out of Scope Issues

• Python 3k will have a much different I/O subsystem. Deciding how that I/O subsystem will work and interact with the bytes object is out of the scope of this PEP. The expectation however is that binary I/O will read and write bytes, while text I/O will read strings. Since the bytes type supports the buffer interface, the existing binary I/O operations in Python 2.6 will support bytes objects.
• It has been suggested that a special method named .__bytes__() be added to the language to allow objects to be converted into byte arrays. This decision is out of scope.
• A bytes literal of the form b"..." is also proposed. This is the subject of PEP 3112.

Open Issues

• The .decode() method is redundant since a bytes object b can also be decoded by calling unicode(b, <encoding>) (in 2.6) or str(b, <encoding>) (in 3.0). Do we need encode/decode methods at all? In a sense the spelling using a constructor is cleaner.
• Need to specify the methods still more carefully.
• Pickling and marshalling support need to be specified.
• Should all those list methods really be implemented?
• A case could be made for supporting .ljust(), .rjust(), .center() with a mandatory second argument.
• A case could be made for supporting .split() with a mandatory argument.
• A case could even be made for supporting .islower(), .isupper(), .isspace(), .isalpha(), .isalnum(), .isdigit() and the corresponding conversions (.lower() etc.), using the ASCII definitions for letters, digits and whitespace. If this is accepted, the cases for .ljust(), .rjust(), .center() and .split() become much stronger, and they should have default arguments as well, using an ASCII space or all ASCII whitespace (for .split()).

Frequently Asked Questions

Q: Why have the optional encoding argument when the encode method of Unicode objects does the same thing?

A: In the current version of Python, the encode method returns a str object and we cannot change that without breaking code. The construct bytes(s.encode(...)) is expensive because it has to copy the byte sequence multiple times. Also, Python generally provides two ways of converting an object of type A into an object of type B: ask an A instance to convert itself to a B, or ask the type B to create a new instance from an A. Depending on what A and B are, both APIs make sense; sometimes reasons of decoupling require that A can’t know about B, in which case you have to use the latter approach; sometimes B can’t know about A, in which case you have to use the former.

Q: Why does bytes ignore the encoding argument if the initializer is a str? (This only applies to 2.6.)

A: There is no sane meaning that the encoding can have in that case. str objects are byte arrays and they know nothing about the encoding of character data they contain. We need to assume that the programmer has provided a str object that already uses the desired encoding. If you need something other than a pure copy of the bytes then you need to first decode the string. For example:

bytes(s.decode(encoding1), encoding2)

Q: Why not have the encoding argument default to Latin-1 (or some other encoding that covers the entire byte range) rather than ASCII?

A: The system default encoding for Python is ASCII. It seems least confusing to use that default. Also, in Py3k, using Latin-1 as the default might not be what users expect. For example, they might prefer a Unicode encoding. Any default will not always work as expected. At least ASCII will complain loudly if you try to encode non-ASCII data.

Copyright

This document has been placed in the public domain.