gh-156743: Document deque.index bound resolution change in 3.14.6

[MannXo]

deque.index() changed how it resolves start and stop when a bound's
__index__ method mutates the deque. The change is user visible and is not
mentioned in the NEWS entry for gh-150750, which describes only the
free-threading race. gh-156743 reports it.

Before the change, Argument Clinic assigned stop its default of
Py_SIZE(deque) before _PyEval_SliceIndexNotNone converted start, so a
start that mutated the deque during conversion left stop holding the old
length while start was resolved against the new one. stop now defaults to
PY_SSIZE_T_MAX and deque_index_impl reads the length once, after both
converters have run, so the two bounds always agree.

Using the reproducer from the issue:

from collections import deque

d = deque([1, 2, 3])

class Grow:
    def __index__(self):
        d.extend([2, 2, 2])
        return -1

d.index(2, Grow())

3.14.5 and earlier raise ValueError. 3.14.6 and later return 5, which is
also what list.index returns for the equivalent list. The new behaviour takes
a single length snapshot and is self-consistent, so it reads as intended and
only the documentation was missing.

This adds a versionchanged note to deque.index. Documentation only, so no
NEWS entry.

One correction to the issue

The issue reports the change as landing in 3.14.7. It landed in 3.14.6.

• The NEWS entry for Data race between deque.index and deque mutators on Py_SIZE with free-threading #150750 is in Misc/NEWS.d/3.14.6.rst.
• Modules/_collectionsmodule.c is identical between the v3.14.6 and
  v3.14.7 tags apart from one unrelated Py_DECREF removal in
  deque_extend_impl.
• The clinic default is Py_SIZE(deque) at v3.14.5 and PY_SSIZE_T_MAX at
  v3.14.6.
• Running the reproducer above on 3.14.6 returns 5.

The versions tested in the issue were 3.14.0, 3.14.2, 3.14.4 and 3.14.7, which
brackets the change but skips 3.14.6. The note therefore says 3.14.6.

[read-the-docs-community]

Documentation build overview

📚 cpython-previews | 🛠️ Build #34649417 | 📁 Comparing 1ce504f against main (1d0f1ad)

  🔍 Preview build  

1 file changed

± library/collections.html

[picnixz]

This kind of changes is not worth documenting. Mutating a container while doing other things is not safe. The fix was just there to prevent crashes I believe.

[MannXo]

Fair, and the fix was for the crash rather than for this. I opened it because JelleZijlstra asked for the note on the issue, after the reporter said he hit it porting pydantic/monty to an alternate implementation rather than fuzzing:

Agree we should document the behavior change.

If the call is not to document it, I will close this, and #156743 wants closing too rather than staying open with a docs label on 3.14, 3.15 and 3.16.

Either way, one correction. The change landed in 3.14.6, not 3.14.7. The clinic default for stop is Py_SIZE(deque) at v3.14.5 and PY_SSIZE_T_MAX at v3.14.6, and the NEWS entry for gh-150750 is in Misc/NEWS.d/3.14.6.rst. The reporter tested 3.14.4 and 3.14.7 and skipped 3.14.6.

[picnixz]

Mmj let me think about it a bit more. I can see why it can help but I do not consider that anything python impmementations should bother about (they can raise whatever they want). It is, IMO, a garbage-in/garbage out situation but let me think for a few days.