[wip/test] ci:Update CI testing matrix for Python versions

[bhimrazy]

What does this PR do ?

Expand the CI testing matrix to include Python versions 3.12 and 3.13.

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 83%. Comparing base (abfebfd) to head (cd0b0d0). Report is 1 commits behind head on main.

Additional details and impacted files

@@         Coverage Diff          @@
##           main   #589    +/-   ##
====================================
+ Coverage    80%    83%    +3%     
====================================
  Files        43     43            
  Lines      6492   6609   +117     
====================================
+ Hits       5181   5499   +318     
+ Misses     1311   1110   -201     

:rocket: New features to boost your workflow:

• :snowflake: Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[deependujha]

is this pr still in WIP? Is there anything else that you intend to add?

[bhimrazy]

is this pr still in WIP? Is there anything else that you intend to add?

Not really. I just had a concern: adding more Python versions might introduce unnecessary tests and increase GitHub Actions usage.

@borda @tchaton, what are your thoughts on this? Should we expand the Python version matrix further, or is the current setup sufficient? I added for Python 3.11 last week.

[bhimrazy]

Tests on both ubuntu-22.04, 3.12 and macos-14, 3.12 appear to have fully passed (292 passed) in ~20 mins and ~30 mins respectively, but the job remains stuck in the "Tests" step and doesn't complete.

Will investigate the cause of the hang shortly.

[Borda]

Seems that the tests are running for 40 minutes, which is quite a significant time, so can we run them in parallel?

[Borda]

The hang on 3.12+ is a known issue in the older pytest-cov/coverage releases: their C-API hooks never return on CPython 3.12+, so pytest never tears down.

seems it did not help

[Borda]

could be related:

Review Python 3.12’s release notes (e.g., PEP 684, multiprocessing improvements) for clues. The spawn start method is safer but slower, which might affect litdata’s performance tests.

[Borda]

@bhimrazy, it seems something in the processing folder is causing hanging, probably not properly cleaned daemon processes... You can try to narrow it down with the ignore argument 🦩

[bhimrazy]

Sure @Borda , will check it out — thanks for the heads-up!

[bhimrazy]

Notes for follow-up PRs:

• Re-enable and fix the skipped test above (test_optimize_checkpoint_in_none_and_append_mode).
• Investigate and Reintroduce macOS for Python ≥3.12
• Add Windows for Python >3.12 to the matrix.