bug: systematic CI failures across PRs and main (torch-nans StopIteration, json sort OOM, io_uring ENOMEM, S3 timeouts)

[u70b3]

Summary

Since ~2026-08-26 09:00 UTC, CI has been failing systematically across independent PRs and main itself in at least five unrelated test families. None of the failures correlate with the PR contents they ran on (verified: the same tests fail on PRs touching disjoint subsystems, and main's own runs are red). This issue consolidates the evidence so the ongoing stabilization work (#8733 → #8745, #8784) can cover the remaining gaps.

Failure families (with run evidence)

1. test_torch_index_with_nans — StopIteration (4+ occurrences, NOT covered by #8784)

Run	Job	Base
32951304277	Python 3.13 (PR #8112)	7b80136
32966813333	Python 3.10 [V3] + [Legacy] (PR #7962)	e49a421
32968165571	Python 3.10 [Legacy] (PR #7962)	e49a421

The test (python/python/tests/test_vector_index.py:548) builds IVF_PQ with num_sub_vectors=4 — the lightweight 4-bit matrix changed in #8733. Torch PQ initialization samples 256 valid residuals from 312 non-NaN rows; StopIteration suggests the filtered sample iterator intermittently exhausts. Failed on PRs touching only mem_wal/api.rs (#8112) and only the scalar cache (#7962) — not PR-caused.

2. test_json_btree_index_null_at_path — DataFusion external sort OOM (2 occurrences)

Run	Job
32968320570	linux-build (PR #7929)
32970551394	linux-build (PR #8112)

ResourcesExhausted: Not enough memory to continue external sort ...
fair(pool_size: 150.0 MB) — Failed to allocate additional 40.0 MB for ExternalSorterMerge[0]

Fails under parallel test load when the memory-pool fair share shrinks. rust/lance-index/src/scalar/json.rs was recently touched by #8577 (skip null tracking for BTree/Bitmap) — worth checking whether the sort input profile changed. Failed on PRs touching only mem-wal — not PR-caused.

3. io_uring worker allocation ENOMEM (2 occurrences)

Run	Test
32951304358	uring::tests::test_read_range (PR #8112)
32957173249	uring::tests::test_read_small_file (PR #7929)

no io_uring worker threads are available: thread 0: Cannot allocate memory (os error 12)

Runner-environment issue (ring allocation restricted on some runners). Same family as #8723 (fixed hang; this is the allocation-failure manifestation).

4. Vector IVF recall/assert flakes (main red + PR runs)

• main Python run 32956317063 @ 4df8219: test_ground_truth, test_ann (assert 11 == 15)
• PR fix(index): replace scalar cache entries on store rotation #7962 mac-build job 98171017852: test_create_ivf_hnsw_pq_multivec
• PR #7929 Rust run 32968320570: test_create_ivf_pq_cosine

Partially addressed by #8784 (4-bit PQ recall) — may already be covered.

5. Environment races (1 occurrence each)

• Windows temp-spill race: test_build_ivf_flat::case_3 — "failed to persist temp file ... The system cannot find the file specified" (PR fix(index): replace scalar cache entries on store rotation #7962 windows-build, run 32951474604)
• S3 network timeout: object_store::tests::test_cloud_paths — reqwest Connect TimedOut resolving bucket.s3.amazonaws.com region (PR fix(index): replace scalar cache entries on store rotation #7962, run 32973042691)

Impact

Every open PR is rolling dice on ~35 jobs × thousands of tests; full-matrix retries (~40 min each) mostly re-fail on a different family. Contributors without write access cannot re-run failed jobs, so PRs sit red even though their code is green. Families 1–3 look deterministically reproducible under load and deserve real fixes (or marked-flaky skips with tracking), not just retries.

2026-08-31 targeted ARM follow-up

Xuanwo directly asked Yang to investigate the Linux ARM/Python 3.14 python/tests/torch_tests/test_bench_utils.py::test_ground_truth failure surfaced on PR #8893. The failing job differs from x86_64 and macOS ARM by one nearest-neighbor index under torch.allclose. First classify equal-distance/tie-order behavior versus floating-point tolerance versus an actual ARM computation defect; do not broaden tolerance until the intended invariant is proven. A Codex worktree setup was requested from clean origin/main@f39a275ccd22d9c1a4d6a95346d0c0cb50d45ade; setup returned only a client thread ID, so this targeted subtask remains Todo until a readable implementation thread exists.

[u70b3]

Claiming family 2 (the spill-pool OOM): #8819 serializes all 15 spill-sort tests in lance-index against the process-wide pool. Root-cause analysis is in the PR description — the short version: sort_spill_reservation_bytes = min(pool/3, 40MB) makes the 150MB FairSpillPool fit exactly 3 concurrent spill sorts, and only the linux-arm job runs cargo test (single shared process), which is why all three observed failures were ARM-only. Families 1 (torch-nans) and 3 (io_uring) are covered by #8767 (pending coverage confirmation on the torch one-pass path) and #8787/#8788 respectively.

[u70b3]

Update: family 1 (torch-nans) is now also covered by #8819 — the gatefixer's reply on #8767 pinpointed next(iter(ds_init)), and the full causal chain turned out to be: a float-NaN vector picked as the IVF init centroid (2.5% with num_partitions=1 and 8/320 NaN rows) → all distances NaN → kmeans "converges" on epoch 0 → every row dropped at partition assignment → empty residuals dataset → PQ init sample yields zero batches → StopIteration. The PR adds _sample_finite_vectors to skip non-finite rows during centroid init (production-path fix, since any NaN-containing vector column + accelerator one-pass IVF-PQ could hit this), with deterministic regression tests. Remaining uncovered families here: io_uring (covered by #8787/#8788) and the one-off env races (Windows spill, S3 timeout).

[wombatu-kun]

Adding one family that is not on the list above: index::vector::ivf::v2::tests::test_legacy_ivf_pq_cosine_multivec_smoke, reporting recall: 0.48 against a 0.5 bar in runs 33125150016, 33108104729 and 33067105290, all on 2026-08-27.

Same root cause as #8764, unseeded k-means centroid initialization, but it builds through pq_matrix_params rather than lightweight_pq_params*, so #8767 does not currently cover it. The two changes it would need on top of that PR are in #8767 (comment).

For anyone weighing a rerun: of the 31 completed, non-cancelled Rust workflow runs on main since 2026-08-25, 16 were green, a 52% pass rate.