homepage

This issue is to add "buffer protocol" (or perhaps "buffer object") to the glossary. The concept is currently described here:
http://docs.python.org/dev/c-api/buffer.html#buffer-protocol
Éric initially suggested doing this in the comments to issue 13538.
Such a glossary entry would be useful because the buffer protocol (or buffer object) should likely be cited, for example, wherever a function accepts a bytes object, bytearray object, or any object that supports the buffer protocol. The str() constructor is one example where this is done:
http://hg.python.org/cpython/file/59acd5cac8b5/Doc/library/functions.rst#l1275
"Buffer object" might be the more useful term to add to the glossary because it would help to have a briefer way of saying "any object that supports the buffer protocol." (I'm assuming this is what "buffer object" actually means.)
The patch for this issue should also do a comprehensive review of occurrences of buffer object/protocol throughout the docs and add or update links and index entries where appropriate.

I would use the term that is currently used in some error messages.

"Buffer protocol" is the right term. "Buffer object" doesn't mean anything in Python 3 and, furthermore, it might be mixed up with the Python 2 `buffer` type.
As for the error messages, they are generally very bad on this topic, so I would vote to change them :-)

Do we have a recommended (and preferably briefer) way of saying, "any object that supports the buffer protocol"?

s/any//

> "Buffer object" doesn't mean anything in Python 3 and, furthermore,
> it might be mixed up with the Python 2 `buffer` type.
Agreed.
> As for the error messages, they are generally very bad on this topic,
> so I would vote to change them :-)
I would say that they are verbose maybe, but not necessary bad.
Using "any object that supports the buffer protocol" without explicitly mentioning bytes (and bytearray) might end up being even more confusing (if that's what it's being proposed).

> Do we have a recommended (and preferably briefer) way of saying, "any
> object that supports the buffer protocol"?
It depends where. There's no recommended way yet, but I would vote for
"bytes-like object" in error messages that are targetted at the average
developer.
The docs (glossary?) could explain that "bytes-like object" is the same
as "buffer-providing object" or "object implementing the buffer
protocol".

> I would vote for "bytes-like object"
Sounds like a good compromise between brevity and clarity to me.

I wouldn't use "bytes-like object". One can certainly argue that *memoryview*
should be bytes-like as a matter of preference, but the buffer protocol
specifies strongly (or even statically) typed multi-dimensional arrays.
PEP-3118 Py_buffer structs are essentially how NumPy works internally.

> I wouldn't use "bytes-like object". One can certainly argue that *memoryview*
> should be bytes-like as a matter of preference, but the buffer protocol
> specifies strongly (or even statically) typed multi-dimensional arrays.
Ach :-(
> PEP-3118 Py_buffer structs are essentially how NumPy works internally.
Well, we should still write a Python documentation, not a NumPy
documentation (on this tracker anyway). Outside of NumPy, there's little
use for multi-dimensional objects.

> I wouldn't use "bytes-like object".
What about "buffer-like object"?

> > I wouldn't use "bytes-like object".
> 
> What about "buffer-like object"?
"buffer-like" means "like a buffer" which is wrong on two points:
- "buffer" is not defined at this point, so the user doesn't understand
what it means
- we are not talking about an object which is "like a buffer", but which
"provides a buffer"

Antoine Pitrou <report@bugs.python.org> wrote:
> > PEP-3118 Py_buffer structs are essentially how NumPy works internally.
> 
> Well, we should still write a Python documentation, not a NumPy
> documentation (on this tracker anyway). Outside of NumPy, there's little
> use for multi-dimensional objects.
Ok, but people should not be surprised if their (Python) array.array() of
double or their array of ctypes structs is silently accepted by some byte
consuming function.
How about "object does not provide a byte buffer" for error messages
and "(byte) buffer provider" as a shorthand for "any buffer provider
that exposes its memory as a sequence of unsigned bytes in response
to a PyBUF_SIMPLE request"?

> > Well, we should still write a Python documentation, not a NumPy
> > documentation (on this tracker anyway). Outside of NumPy, there's little
> > use for multi-dimensional objects.
> 
> Ok, but people should not be surprised if their (Python) array.array() of
> double or their array of ctypes structs is silently accepted by some byte
> consuming function.
Probably. My own (humble :-)) opinion is that array.array() is a
historical artifact, and its use doesn't seem to be warranted in modern
Python code. ctypes is obviously a very special library, and not for the
faint of heart.
> How about "object does not provide a byte buffer" for error messages
> and "(byte) buffer provider" as a shorthand for "any buffer provider
> that exposes its memory as a sequence of unsigned bytes in response
> to a PyBUF_SIMPLE request"?
It's not too bad, I think. However, what I think is important is that
the average (non-expert) Python developer understand that the function
really accepts a bytes object, and other similar types (because, really,
bytes is the only bytes-like type most developers will ever face).
That's why I'm proposing "bytes-like object".

Antoine Pitrou <report@bugs.python.org> wrote:
> > How about "object does not provide a byte buffer" for error messages
> > and "(byte) buffer provider" as a shorthand for "any buffer provider
> > that exposes its memory as a sequence of unsigned bytes in response
> > to a PyBUF_SIMPLE request"?
> 
> It's not too bad, I think. However, what I think is important is that
> the average (non-expert) Python developer understand that the function
> really accepts a bytes object, and other similar types (because, really,
> bytes is the only bytes-like type most developers will ever face).
> That's why I'm proposing "bytes-like object".
If it is somehow possible to establish the term as a shorthand for the real
meaning, then I guess it's the most economical option for documenting Python
methods (I don't think it should be used in the C-API docs though).
help (b''.join) for example would sound better with "bytes-like object"
than with "(byte) buffer provider".

> > That's why I'm proposing "bytes-like object".
>
> If it is somehow possible to establish the term as a shorthand for the real
meaning,
This can be established via the glossary. We can still use "buffer provider" for the general case, if we find that it is useful in certain circumstances.

After this issue is resolved, the binascii docs can be updated as suggested in issue 16724.

Here's a patch that adds "bytes-like object" to the glossary, links to the buffer protocol docs[0] and provides bytes and bytearray as examples.
[0]: http://docs.python.org/dev/c-api/buffer.html#buffer-protocol 

New changeset 474f28bf67b3 by Ezio Melotti in branch '3.3':
#16518: add "bytes-like object" to the glossary.
http://hg.python.org/cpython/rev/474f28bf67b3
New changeset 747cede24367 by Ezio Melotti in branch 'default':
#16518: merge with 3.3.
http://hg.python.org/cpython/rev/747cede24367
New changeset 1b92a0112f5d by Ezio Melotti in branch '2.7':
#16518: add "bytes-like object" to the glossary.
http://hg.python.org/cpython/rev/1b92a0112f5d 

New changeset d1aa8a9eba44 by Ezio Melotti in branch '2.7':
#16518: fix links in glossary entry.
http://hg.python.org/cpython/rev/d1aa8a9eba44 

The attached patch replaces things like "object that support the buffer protocol/interface/API" with "bytes-like objects" throughout the docs.
The patch doesn't change error messages/docstrings.
I also noticed that on 2.7[0], the section about the buffer protocol in Doc/c-api/buffer.rst is called "Buffers and Memoryview Objects" and it's not as clear as the one on 3.x[1]. Should this section be backported?
[0]: http://docs.python.org/2.7/c-api/buffer.html#bufferobjects
[1]: http://docs.python.org/dev/c-api/buffer.html#bufferobjects 

> I also noticed that on 2.7[0], the section about the buffer protocol
> in Doc/c-api/buffer.rst is called "Buffers and Memoryview Objects" and
> it's not as clear as the one on 3.x[1]. Should this section be
> backported?
The "buffer protocol" situation is different on 2.x, please let's
concentrate on 3.x :-)

New changeset 003e4eb92683 by Ezio Melotti in branch '3.3':
#16518: use "bytes-like object" throughout the docs.
http://hg.python.org/cpython/rev/003e4eb92683
New changeset d4912244cce6 by Ezio Melotti in branch 'default':
#16518: merge with 3.3.
http://hg.python.org/cpython/rev/d4912244cce6 

The attached patch uses "bytes-like objects" in the error messages.

> The attached patch uses "bytes-like objects" in the error messages.
I'm surprised your patch doesn't touch Python/getargs.c.

FWIW I was grepping for buffer protocol/interface/api, and then double-checking for "buffer" in the resulting files. Python/getargs.c doesn't seem to mention the buffer protocol/interface/api at all.

Updated patch to include getargs.c too.

At first-reading, it looks like matters were made more confusing with "bytes-like object" as a defined term.

Can you elaborate?

New changeset e7e8a218737a by R David Murray in branch 'default':
#16518: Bring error messages in harmony with docs ("bytes-like object")
https://hg.python.org/cpython/rev/e7e8a218737a 

Committed the message changes to 3.5 only, since it will probably cause tests to fail in various projects, despite messages not being a formal part of the python API. 
Per IRC conversation with Ezio and Antoine, I posted a note to python-dev to let people know we now have a consistent terminology in the docs and error messages, and to provide a last opportunity for objections (it is easy enough to back the patch out if there is an outcry, but I don't expect one).

There are other unfixed messages (may be introduced after 3.3):
>>> b''.join([''])
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
TypeError: sequence item 0: expected bytes, bytearray, or an object with the buffer interface, str found
>>> str(42, 'utf8')
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
TypeError: coercing to str: need bytes, bytearray or buffer-like object, int found
>>> import array; array.array('B').frombytes(array.array('I'))
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
TypeError: string/buffer of bytes required.
>>> import socket; print(socket.socket.sendmsg.__doc__)
sendmsg(buffers[, ancdata[, flags[, address]]]) -> count
Send normal and ancillary data to the socket, gathering the
non-ancillary data from a series of buffers and concatenating it into
a single message. The buffers argument specifies the non-ancillary
data as an iterable of buffer-compatible objects (e.g. bytes objects).
The ancdata argument specifies the ancillary data (control messages)
as an iterable of zero or more tuples (cmsg_level, cmsg_type,
cmsg_data), where cmsg_level and cmsg_type are integers specifying the
protocol level and protocol-specific type respectively, and cmsg_data
is a buffer-compatible object holding the associated data. The flags
argument defaults to 0 and has the same meaning as for send(). If
address is supplied and not None, it sets a destination address for
the message. The return value is the number of bytes of non-ancillary
data sent.
And there are several mentions of "buffer-like" or "buffer-compatible" in the documentation.

Please open a new issue for those.

See #22581.