test: harden Redis-backed test isolation under pytest-xdist (#546)

[vishal-bala]

Motivation

The integration suite runs under pytest -n auto, and now and then a single matrix job fails on a test that passes everywhere else — and passes again on rerun. The most recent example was test_hybrid_query_with_geo_filter returning 0 results instead of 3 on exactly one Python/redis-py/Redis combination. These failures read as version-specific, but they aren't: they're a test-isolation problem wearing a disguise.

The key fact is how the suite isolates workers. Each xdist worker gets its own Redis container, so two tests can only interfere when they run on the same worker and reach for the same Redis key names. For that to actually corrupt a result, two things have to line up:

1. a shared resource name (a name scoped only to the worker, or a hard-coded one), and
2. a way for state to outlive its test — cleanup that only runs on the success path, or an index recreated with overwrite=True but no drop=True, leaving stale documents under a reused prefix.

When both hold, one test sees another's leftovers — empty results, inflated counts, dtype mismatches — and which test loses depends entirely on the order the scheduler picked. That's the retry-passing, pseudo-random signature.

#543 introduced a redis_test_name helper that mints per-test-unique names and began migrating fixtures onto it. This PR finishes that migration for the modules that can genuinely cause flakiness, and tightens their recreate/cleanup so leftovers can't survive.

What changed

Recreate group — per-test names + create(overwrite=True, drop=True):

• test_hybrid.py and test_aggregation.py (these two shared the identical user_index_{worker_id} / v1_{worker_id} — the direct cause of the geo-filter failure), plus test_search_results.py, test_svs_integration.py, and the shared conftest.py index fixtures.

Fixed-name collisions — per-test names + guaranteed teardown:

• from_existing_complex (hard-coded "test") and no_proactive_module_validation ("my_index") in the sync + async search-index suites.
• The three test_embedcache_warnings.py tests sharing "test_cache".
• The shared "doc" prefix / "test-index" name across test_no_proactive_module_checks.py.

Raw worker_id collision — unique names + cleanup:

• The two test_llmcache.py tests that both created float64_cache_{worker_id} and never cleaned up.

Cluster tests — the one place cross-worker collisions are real, since every cluster test points at a single hard-coded redis://localhost:7001: test_redis_cluster_support.py and test_cluster_pipelining.py now use per-test names, drop=True, and try/finally cleanup.

Scope and follow-ups

• The SemanticRouter.delete() leak of its {name}:route_config key is a product-code gap, not a test-only fix, so it's split into SemanticRouter.delete() leaves orphaned {name}:route_config key in Redis #634 (which also owns the dependent test_key_separator_handling.py cleanup).
• The remaining Complete Redis-backed test isolation hardening for parallel test runs #546 items are benign hygiene — leaks under per-test-unique names that can't contaminate another test, only accumulate harmless keys in a worker's throwaway Redis. Their exact remaining scope is enumerated in a comment on Complete Redis-backed test isolation hardening for parallel test runs #546.

Testing

Honest caveat for reviewers: this was developed in an environment without Docker, so I could not run the integration suite. black, isort, and mypy pass. Please run make test and a repeated pytest -n auto pass to confirm the flake is gone before merging — this is a test-only change, so behavior is fully validated by the suite itself.

---

Note

Low Risk
Changes are confined to the test suite; production library behavior is unchanged.

Overview
Hardens integration tests so parallel pytest-xdist workers and shared Redis/cluster endpoints do not leave stale keys or collide on index/cache names.

Fixtures and tests now use redis_test_name (per-test unique index/prefix/cache names) instead of worker_id-only or hard-coded names like "test" / "my_index" / "test_cache". Shared conftest index fixtures (flat_index, hnsw_index, async variants) and modules such as test_hybrid, test_aggregation, test_search_results, and test_svs_integration adopt that naming and call create(overwrite=True, drop=True) so documents under a reused prefix are cleared before load.

Teardown is tightened with try/finally, delete(drop=True), and cache clear() / aclear() on failure paths (e.g. from_existing_complex, module-validation tests, test_llmcache dtype caches, test_embedcache_warnings). Cluster tests use per-test names and drop=True because all workers share localhost:7001.

^{Reviewed by Cursor Bugbot for commit c4e6b33. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jit-ci]

🛡️ Jit Security Scan Results

✅ No security findings were detected in this PR

---

^{Security scan by Jit}