The tempfile module contains the class _RandomNameSequence, which has the rng property.
This property checks os.getpid() every time and re-initializes a random number generator when it has changed.
However, this is only necessary on systems which allow the process to be cloned and should be solved on such systems with os.register_at_fork (see the random module).

Wouldn't it be simpler to use random.SystemRandom instead?

SystemRandom seems to be slower:


from random import Random, SystemRandom
from timeit import timeit

user = Random()
system = SystemRandom()
characters = "abcdefghijklmnopqrstuvwxyz0123456789_"

timeit(lambda: user.choice(characters))
>>> 0.5491522020020057

timeit(lambda: system.choice(characters))
>>> 2.9195130389998667

I also wonder if the overhead of getpid() is actually significant for anything.

It seems to be insignificant, however it would allow for easier monkey-patching: https://bugs.python.org/issue32276

Instead of changing _Random one could simply assign a new instance to _named_sequence

New changeset 8e409cebad42032bb7d0f2cadd8b1e36081d98af by Eric W in branch 'master':
bpo-42160: tempfile: Reduce overhead of pid check. (GH-22997)
https://github.com/python/cpython/commit/8e409cebad42032bb7d0f2cadd8b1e36081d98af

Thanks

Thanks

The commit 8e409cebad42032bb7d0f2cadd8b1e36081d98af introduced a reference leak:
https://buildbot.python.org/all/#/builders/320/builds/71

$ make && ./python -m test test_tempfile -R 3:3
...
test_tempfile leaked [49, 49, 49] references, sum=147
test_tempfile leaked [28, 28, 28] memory blocks, sum=84
...

Single test reproducing the leak:

$ ./python -m test test_tempfile -R 3:3 -m test.test_tempfile.TestGetDefaultTempdir.test_no_files_left_behind
...
test_tempfile leaked [21, 21, 21] references, sum=63
test_tempfile leaked [12, 12, 12] memory blocks, sum=36
...

Oh, _RandomNameSequence is not true singleton. Instance is used in tests actually.
Using `os.register_at_fork` was but idea. Let's reject it.

New changeset 43ca084c88d1e46a44199f347c72e26db84903c9 by Inada Naoki in branch 'master':
Revert "bpo-42160: tempfile: Reduce overhead of pid check. (GH-22997)"
https://github.com/python/cpython/commit/43ca084c88d1e46a44199f347c72e26db84903c9

@Deric-W Please create another PR to change from `choise` to `choises` if you want to optimize it.

See also issue30030.

It would be possible to allow the GC to finalize the Random instances through weak references.
On the other hand, if many _RandomNameSequence instances were used temporarily, a lot of callbacks would be registered via os.register_at_fork(), could that cause problems?

Instead of os.getpid() we can use a global variable, and the single callback can reset it.

But is it worth? Is os.getpid() so slow?

>>> timeit(os.getpid)
0.08990733299999931

Considering the reference leaks, os.getpid() seems to be the better solution.

I have tested my idea, such optimization speeds up _RandomNameSequence by 16%. At the cost of more complex code. It is not worth.