Speed up high-impact functional tests and replace slow ones with unit tests

[gerrod3]

Summary

• Speed up the functional tests that showed the largest CI wins (shared setup for replication, pulpimport, and pulp export; pagination; MIME types; checkpoints; hidden distros; cross-domain exports).
• Replace the Redis fetch-batch functional test (60s sleep, 25 tasks) with a unit test, and add a unit test for missing-worker cleanup so the long-running functional test can stay nightly-only.
• Timed 5× before/after on isolated lean stacks. Biggest saves: replication group ~127s (1.56x), pulpimport group ~62s (2.64x), repo-version pagination ~45s (5.76x).

Test plan

• oci-env test -p pulpcore unit -- -k "test_missing_worker_cleanup_fails_abandoned_task or test_fetch_task_beyond_initial_batch"
• oci-env test -p pulpcore functional -- -k "TestPulpImport or test_repo_version_pagination or TestCheckpointDistribution or test_checkpoint_retention or test_hidden_distros"
• oci-env test -p pulp_file functional -- -k "test_content_types or TestSyncedRepoExport or test_fsexport_cross_domain"
• Replication/cross-domain tests with PULP_DOMAIN_ENABLED=True

Made with Cursor

[dralley]

Is this officially replacing #7968? If so can that PR be closed?

[gerrod3]

Is this officially replacing #7968? If so can that PR be closed?

Yes, once I'm done comparing the final times and ensuring the most impactful have been extracted, I'll close the old one.

[gerrod3]

The speedup is kind of insane, a sub-20 azure run... Not sure if I just got lucky with the runner, but here's the explanation from AI:

CI is faster because the old tests were paying for setup over and over, not because the assertions got cheaper.

On test (pulp), pytest time vs a recent full-suite PR (fix/7910):

Suite	Before	This PR	Saved
pulpcore functional parallel (-n 8)	6:33 (212 tests)	4:00 (192)	2:33
pulpcore functional serial	2:33 (36 tests)	2:24 (57)	~9s with 21 extra tests
pulp_file functional parallel	3:31 (99)	2:32 (95)	59s
pulp_file functional serial	25s (3)	34s (7)	+9s (export class moved here)

Job wall clock: pulp 21 → 16 min, azure 26 → 18 min, s3 25 → 21 min. Azure dropped more because that matrix runs domains, where the replication suite was the biggest local win (~353s → ~226s).

Three mechanisms show up in the logs:

1. Stop repeating expensive setup under xdist.
Pulpimport and pulp export used to sync/export on every test. With -n 8, that setup ran many times at once and contended for the worker. They are now a serial class with class-scoped fixtures, so the two immediate syncs / three-repo export happen once. Parallel count dropped ~20 tests; serial picked them up and still got faster. Checkpoint did the same: five publications used to be rebuilt on every worker.

2. Skip the task round-trip for content create.
ContentFilesApi.create + monitor_task is a full task. upload (or overlapping creates) makes the unit appear in-process. That is most of pagination (55s → 9s locally), MIME types, checkpoints, and file_9_contents.

3. Do less work that the assertion does not need.
Replication used a 3-file fixture sync per upstream; it now uploads one file. Hidden distros created ten distributions; four still covers the index. Pagination used a large page size, so it had to create that many versions. The Redis fetch-batch test slept 60s and queued 25 tasks; that is a unit test now.

The isolated benches understated CI impact: a single pytest -k cannot show “setup once, eight workers.” CI with -n 8 is exactly where shared fixtures pay off.