I went down this rabbit hole when someone mentioned that isfile/isdir/exists all make a rather expensive os.stat call on Windows (which is actually a long wrapper around a number of system calls on Windows), rather than the simpler and more direct call to GetFileAttributeW.

I noticed that at one point there was a version of isdir that does exactly this. At the time, this claimed a 2x speedup.

However, this C implementation of isdir was removed as part of a large set of changes in df2d4a6, and as a result, isdir got faster.

With the following benchmark:

import os.path
import timeit


for i in range(100):
    os.makedirs(f"exists{i}", exist_ok=True)


def test_exists():
    for i in range(100):
        os.path.isdir(f"exists{i}")


def test_extinct():
    for i in range(100):
        os.path.isdir(f"extinct{i}")


print(timeit.timeit(test_exists, number=100))
print(timeit.timeit(test_extinct, number=100))


for i in range(100):
    os.rmdir(f"exists{i}")

I get the following with df2d4a6:

exists: 0.18694799999957468
doesn't exist: 0.08418370000072173

and with the prior commit:

exists: 0.25393609999991895
doesn't exist: 0.08511730000009265

So, from this, I'd conclude that the idea of replacing calls to os.stat with calls to GetFileAttributeW would not bear fruit, but @zooba should probably confirm I'm benchmarking the right thing and making sense.

In any event, we should probably remove the little vestige that imports this fast path that was removed:

try:
    # The genericpath.isdir implementation uses os.stat and checks the mode
    # attribute to tell whether or not the path is a directory.
    # This is overkill on Windows - just pass the path to GetFileAttributes
    # and check the attribute from there.
    from nt import _isdir as isdir
except ImportError:
    # Use genericpath.isdir as imported above.
    pass

Linked PRs

• gh-101196: Make isdir/isfile/exists faster on Windows #101324