|Title:||Immutable Bytes and Mutable Buffer|
|Author:||Guido van Rossum <guido at python.org>|
After releasing Python 3.0a1 with a mutable bytes type, pressure mounted to add a way to represent immutable bytes. Gregory P. Smith proposed a patch that would allow making a bytes object temporarily immutable by requesting that the data be locked using the new buffer API from PEP 3118. This did not seem the right approach to me.
Jeffrey Yasskin, with the help of Adam Hupp, then prepared a patch to make the bytes type immutable (by crudely removing all mutating APIs) and fix the fall-out in the test suite. This showed that there aren't all that many places that depend on the mutability of bytes, with the exception of code that builds up a return value from small pieces.
Thinking through the consequences, and noticing that using the array module as an ersatz mutable bytes type is far from ideal, and recalling a proposal put forward earlier by Talin, I floated the suggestion to have both a mutable and an immutable bytes type. (This had been brought up before, but until seeing the evidence of Jeffrey's patch I wasn't open to the suggestion.)
Moreover, a possible implementation strategy became clear: use the old PyString implementation, stripped down to remove locale support and implicit conversions to/from Unicode, for the immutable bytes type, and keep the new PyBytes implementation as the mutable bytes type.
The ensuing discussion made it clear that the idea is welcome but needs to be specified more precisely. Hence this PEP.
One advantage of having an immutable bytes type is that code objects can use these. It also makes it possible to efficiently create hash tables using bytes for keys; this may be useful when parsing protocols like HTTP or SMTP which are based on bytes representing text.
Porting code that manipulates binary data (or encoded text) in Python 2.x will be easier using the new design than using the original 3.0 design with mutable bytes; simply replace str with bytes and change '...' literals into b'...' literals.
I propose the following type names at the Python level:
- bytes is an immutable array of bytes (PyString)
- bytearray is a mutable array of bytes (PyBytes)
- memoryview is a bytes view on another object (PyMemory)
The old type named buffer is so similar to the new type memoryview, introduce by PEP 3118, that it is redundant. The rest of this PEP doesn't discuss the functionality of memoryview; it is just mentioned here to justify getting rid of the old buffer type. (An earlier version of this PEP proposed buffer as the new name for PyBytes; in the end this name was deemed to confusing given the many other uses of the word buffer.)
While eventually it makes sense to change the C API names, this PEP maintains the old C API names, which should be familiar to all.
Here's a simple ASCII-art table summarizing the type names in various Python versions:
+--------------+-------------+------------+--------------------------+ | C name | 2.x repr | 3.0a1 repr | 3.0a2 repr | +--------------+-------------+------------+--------------------------+ | PyUnicode | unicode u'' | str '' | str '' | | PyString | str '' | str8 s'' | bytes b'' | | PyBytes | N/A | bytes b'' | bytearray bytearray(b'') | | PyBuffer | buffer | buffer | N/A | | PyMemoryView | N/A | memoryview | memoryview <...> | +--------------+-------------+------------+--------------------------+
The b'...' notation introduced in Python 3.0a1 returns an immutable bytes object, whatever variation is used. To create a mutable array of bytes, use bytearray(b'...') or bytearray([...]). The latter form takes a list of integers in range(256).
PEP 3118 Buffer API
Both bytes and bytearray implement the PEP 3118 buffer API. The bytes type only implements read-only requests; the bytearray type allows writable and data-locked requests as well. The element data type is always 'B' (i.e. unsigned byte).
There are four forms of constructors, applicable to both bytes and bytearray:
- bytes(<bytes>), bytes(<bytearray>), bytearray(<bytes>), bytearray(<bytearray>): simple copying constructors, with the note that bytes(<bytes>) might return its (immutable) argument, but bytearray(<bytearray>) always makes a copy.
- bytes(<str>, <encoding>[, <errors>]), bytearray(<str>, <encoding>[, <errors>]): encode a text string. Note that the str.encode() method returns an immutable bytes object. The <encoding> argument is mandatory; <errors> is optional. <encoding> and <errrors>, if given, must be str instances.
- bytes(<memory view>), bytearray(<memory view>): construct a bytes or bytearray object from anything that implements the PEP 3118 buffer API.
- bytes(<iterable of ints>), bytearray(<iterable of ints>): construct a bytes or bytearray object from a stream of integers in range(256).
- bytes(<int>), bytearray(<int>): construct a zero-initialized bytes or bytearray object of a given length.
The bytes and bytearray types are comparable with each other and orderable, so that e.g. b'abc' == bytearray(b'abc') < b'abd'.
Comparing either type to a str object for equality returns False regardless of the contents of either operand. Ordering comparisons with str raise TypeError. This is all conformant to the standard rules for comparison and ordering between objects of incompatible types.
(Note: in Python 3.0a1, comparing a bytes instance with a str instance would raise TypeError, on the premise that this would catch the occasional mistake quicker, especially in code ported from Python 2.x. However, a long discussion on the python-3000 list pointed out so many problems with this that it is clearly a bad idea, to be rolled back in 3.0a2 regardless of the fate of the rest of this PEP.)
Slicing a bytes object returns a bytes object. Slicing a bytearray object returns a bytearray object.
Slice assignment to a bytearray object accepts anything that implements the PEP 3118 buffer API, or an iterable of integers in range(256).
Indexing bytes and bytearray returns small ints (like the bytes type in 3.0a1, and like lists or array.array('B')).
Assignment to an item of a bytearray object accepts an int in range(256). (To assign from a bytes sequence, use a slice assignment.)
The str() and repr() functions return the same thing for these objects. The repr() of a bytes object returns a b'...' style literal. The repr() of a bytearray returns a string of the form "bytearray(b'...')".
The following operators are implemented by the bytes and bytearray types, except where mentioned:
- b1 + b2: concatenation. With mixed bytes/bytearray operands, the return type is that of the first argument (this seems arbitrary until you consider how += works).
- b1 += b2: mutates b1 if it is a bytearray object.
- b * n, n * b: repetition; n must be an integer.
- b *= n: mutates b if it is a bytearray object.
- b1 in b2, b1 not in b2: substring test; b1 can be any object implementing the PEP 3118 buffer API.
- i in b, i not in b: single-byte membership test; i must be an integer (if it is a length-1 bytes array, it is considered to be a substring test, with the same outcome).
- len(b): the number of bytes.
- hash(b): the hash value; only implemented by the bytes type.
Note that the % operator is not implemented. It does not appear worth the complexity.
The following methods are implemented by bytes as well as bytearray, with similar semantics. They accept anything that implements the PEP 3118 buffer API for bytes arguments, and return the same type as the object whose method is called ("self"):
.capitalize(), .center(), .count(), .decode(), .endswith(), .expandtabs(), .find(), .index(), .isalnum(), .isalpha(), .isdigit(), .islower(), .isspace(), .istitle(), .isupper(), .join(), .ljust(), .lower(), .lstrip(), .partition(), .replace(), .rfind(), .rindex(), .rjust(), .rpartition(), .rsplit(), .rstrip(), .split(), .splitlines(), .startswith(), .strip(), .swapcase(), .title(), .translate(), .upper(), .zfill()
This is exactly the set of methods present on the str type in Python 2.x, with the exclusion of .encode(). The signatures and semantics are the same too. However, whenever character classes like letter, whitespace, lower case are used, the ASCII definitions of these classes are used. (The Python 2.x str type uses the definitions from the current locale, settable through the locale module.) The .encode() method is left out because of the more strict definitions of encoding and decoding in Python 3000: encoding always takes a Unicode string and returns a bytes sequence, and decoding always takes a bytes sequence and returns a Unicode string.
In addition, both types implement the class method .fromhex(), which constructs an object from a string containing hexadecimal values (with or without spaces between the bytes).
The bytearray type implements these additional methods from the MutableSequence ABC (see PEP 3119):
.extend(), .insert(), .append(), .reverse(), .pop(), .remove().
Like the bytes type in Python 3.0a1, and unlike the relationship between str and unicode in Python 2.x, attempts to mix bytes (or bytearray) objects and str objects without specifying an encoding will raise a TypeError exception. (However, comparing bytes/bytearray and str objects for equality will simply return False; see the section on Comparisons above.)
Conversions between bytes or bytearray objects and str objects must always be explicit, using an encoding. There are two equivalent APIs: str(b, <encoding>[, <errors>]) is equivalent to b.decode(<encoding>[, <errors>]), and bytes(s, <encoding>[, <errors>]) is equivalent to s.encode(<encoding>[, <errors>]).
There is one exception: we can convert from bytes (or bytearray) to str without specifying an encoding by writing str(b). This produces the same result as repr(b). This exception is necessary because of the general promise that any object can be printed, and printing is just a special case of conversion to str. There is however no promise that printing a bytes object interprets the individual bytes as characters (unlike in Python 2.x).
The str type currently implements the PEP 3118 buffer API. While this is perhaps occasionally convenient, it is also potentially confusing, because the bytes accessed via the buffer API represent a platform-depending encoding: depending on the platform byte order and a compile-time configuration option, the encoding could be UTF-16-BE, UTF-16-LE, UTF-32-BE, or UTF-32-LE. Worse, a different implementation of the str type might completely change the bytes representation, e.g. to UTF-8, or even make it impossible to access the data as a contiguous array of bytes at all. Therefore, the PEP 3118 buffer API will be removed from the str type.
The basestring type will be removed from the language. Code that used to say isinstance(x, basestring) should be changed to use isinstance(x, str) instead.
Left as an exercise for the reader.
This document has been placed in the public domain.