fix(index): resolve frag-reuse index single-flight self-deadlock in open_frag_reuse_index

[zehiko]

Port of #41 (release-8.0.0) to release-9.0.0.

Mechanism

DatasetIndexInternalExt::open_frag_reuse_index passes a loader closure to index_cache.get_or_insert_with_key(frag_reuse_key, ...) that calls self.load_index(&frag_reuse_uuid). load_index runs load_indices(), which does its own get_or_insert_with_key on the same frag_reuse/{uuid} cache key. moka's optionally_get_with holds the waiter's write lock across the init future, so the re-entrant inner call parks on read().await against its own outer write guard: a permanent, zero-CPU self-deadlock. Every subsequent reader of the dataset then parks behind the poisoned key — all reads of the dataset stall indefinitely.

Trigger conditions

• The frag_reuse/{uuid} entry must be absent at the second get_or_insert_with_key — i.e. evicted in the window between the pre-populate in load_indices() and the probe in open_frag_reuse_index.
• In practice: a huge materialized FRI (large weight → prime eviction target) + FRI uuid churn from deferred-remap compactions + concurrent readers keeping the cache under pressure.

Impact

Root cause of the ab-repro-0155 total registration stall on 2026-08-18, and of RR-5448's previously unexplained "second stall mechanism" on ab-repro-new.

Fix

The loader now uses the IndexMetadata already fetched by load_index_by_name at the top of the function and calls load_frag_reuse_index_details + open_frag_reuse_index directly — mirroring the (safe) loader inside load_indices(). Guard comments added on both FRI loader closures: they must never call load_indices/load_index, which re-enter the same cache key.

Verification

• Deterministic regression test (rust/lance/tests/frag_reuse_deadlock.rs): a zero-capacity index cache forces the poisoned loader path; unfixed = deadlock caught by a 20s tokio timeout, fixed = passes.
• Live verification on ab-repro-0155: the unfixed binary wedges in ~8 min under 80k-registration stress; the fixed binary passed 67 FRI evictions / 14 compactions in 4 min with zero stuck cache waits.

Upstream

Upstream lancedb/lance main carries the same defect (rust/lance/src/index.rs:3003-3005, loader calls self.load_index(...)), as does our release-9.0.0 branch (rust/lance/src/index.rs:2494-2496) — this fix should be forward-ported.

🤖 Generated with Claude Code