homepage

The fact that "buffering=1" is usable only in text mode is documented for open(). In binary mode, setting buffering to 1 is silently ignored and equivalent to using default buffer size. I argue that such behavior is:
1. Inconsistent
For example, attempts to use buffering=0 in text mode or to specify encoding in binary mode raise ValueError.
2. Dangerous
Consider the following code fragment running on *nix (inspired by real-world code):
with open("fifo", "wb", buffering=1) as out:
 for line in lines:
 out.write(line)
"fifo" refers to a FIFO (named pipe). For each line, len(line) <= PIPE_BUF. Because of line buffering, such code must be able to safely assume that each write() is atomic, so that even in case of multiple writers no line read by a FIFO reader will ever get intermixed with another line. And it's so in Python 2. After migration to Python 3 (assuming that line is now bytes), this code still works on Linux because of two factors:
a) PIPE_BUF is 4096, and so is the default IO buffer size (usually)
b) When a write() is going to overflow the buffer, BufferedWriter first flushes and then processes the new write() (based on my experiment). So, each OS write() is called with complete lines only and is atomic per (a).
But PIPE_BUF is 512 on Mac OS X (I don't know about default buffer size). So, we are likely to have a subtle 2-to-3 migration issue.
I suggest to raise a ValueError if buffering=1 is used for binary mode. Note that issue 10344 (msg244354) and issue 21332 would have been uncovered earlier if it was done from the beginning.

> I suggest to raise a ValueError if buffering=1 is used for binary mode.
Either that, or we instead accept buffering=1 as a regular buffer size. But since buffering=1 means something else for text mode, maybe you're right that it's better to raise in binary mode, to avoid any possible confusion.

I'm in favor of raising an exception because it'll expose existing code with incorrect assumptions. I'll check whether it breaks any tests and submit a PR.

After looking at the PR, I think it would be a bit too strong to raise an error. Perhaps emit a warning instead?

I had similar thoughts when I was fixing tests that broke due to ValueError. I've updated the PR to issue a RuntimeWarning instead.

Any feedback on the updated PR?

My problem with a warning is the standard one: People who see a warning are often end users of python applications (who don't even have to know what Python is, let alone know anything about the code). For that reason, never add a warning to a stable branch - only new releases (meaning 3.8 here).
Given that this isn't not a deprecation of meaningful API behavior but is highlighting questionably undefined nonsense behavior, users complaining upon obtaining 3.8 should ultimately reach library and application developers who use the API wrong to update their call sites to explicitly ask for what they intended instead of being ambiguious.
FYI - the subprocess.py related changes in your PR are correct.

Thank you, Gregory. I didn't intend to add the warning to stable branches -- it's just that 3.7 hasn't been released yet when this report was submitted.

New changeset a2670565d8f5c502388378aba1fe73023fd8c8d4 by Victor Stinner (Alexey Izbyshev) in branch 'master':
bpo-32236: open() emits RuntimeWarning if buffering=1 for binary mode (GH-4842)
https://github.com/python/cpython/commit/a2670565d8f5c502388378aba1fe73023fd8c8d4

I don't think that it would be a good idea to start emitting a new warning in a minor release like the future Python 3.7.2, so I suggest to not backport the change.
I close the issue.