[Python] Reading a zero-size buffer can segfault

[asfimport]

Simplest reproducible code is:

pa.read_message(b'')

which gives a segfault.

You can easily run into this interactively when eg by accident passing a already-read buffer to it, like:

serialized = pa.schema([('a', pa.int64())]).serialize().to_pybytes()
buffer = pa.BufferReader(serialized)
pa.read_message(buffer)
pa.read_message(buffer)

And for example, if you compare to read_schema, this gives an error on the second time / empty buffer:

>>> pa.read_schema(buffer)
>>> pa.read_schema(buffer)
...
ArrowInvalid: Tried reading schema message, was null or length 0

I know this is not proper usage of Buffer(Reader), but since it is easy to accidentally do this, we should try to protect users from this I think.

Reporter: Joris Van den Bossche / @jorisvandenbossche
Assignee: Antoine Pitrou / @pitrou

PRs and other links:

• GitHub Pull Request #5437

_{Note: This issue was originally created as ARROW-6618. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
This deserves fixing indeed.

[asfimport]

Wes McKinney / @wesm:
Issue resolved by pull request 5437
#5437