"w" format is dangerous because it doesn't give the size of the buffer: the caller may write outside the buffer (buffer overflow).

"w*" and "w#" formats are fine.

It looks like "w" format is not used in trunk nor py3k (only w# and w*).

Attached patch removes "w" format, cleanups the code for "w*" and "w#" formats, and update the documentation.

I'd point out that "w#" is mostly useless too. It's supposed to return a read-write buffer, but as the doc says it also doesn't support "mutable objects", since it isn't able to properly lock/pin the buffer; therefore it probably doesn't support anything useful.

See also #8592: 'y' does not check for embedded NUL bytes.

"w#" is not only useless but also not used in py3k source code. "w" is also not used. Only "w*" is used by fnctl and socket modules.

The problem with w# is that the caller cannot "release" the buffer and so we cannot lock the buffer. I don't know exactly what happens if you get a pointer to a bytearray, release the GIL and then use it, whereas another thread modifies and/or destroys the bytearray object. (I suppose that something bad will happen, like a segfault)

New (improved) version of the patch:
 - Remove also "w#" format
 - Write tests for "w*" format

See also #8926.

Commited to 3.2 (r82208), blocked in 3.1 (r82209).