Enforce the dropTable cross-worker-write reproducer; close #1381, #2496

[kriszyp]

Summary

Confirms the cross-worker write vs RocksDB table-drop race (#1381, #2496 — "Invalid column family specified in write batch" out of completeInterruptedDrops) is fixed on main, and turns its reproducer into an unconditional, enforced regression test.

unitTests/resources/dropTableCrossWorkerWrite.test.js's second test opened with an assert.ok tripwire on columnFamily.pendingReclaims, a stat that only exists in a @harperfast/rocksdb-js build that defers physical column-family drops behind admitted commits (rocksdb-js#850). It started as an unconditional it.skip, became a conditional this.skip(), then a hard assert so an older binding fails loud instead of skipping silently.

package.json now pins @harperfast/rocksdb-js@2.10.0 (ships #850), and harper's generation-distinct table stores (66c61089d, refs #1381) landed the other half of the fix. The stat always exists at this pin, so the assertion was a permanent no-op. This PR removes it so the reproducer runs unconditionally, and adds a one-line comment pointing at rocksdb-js#850 in its place — the deleted assertion was the file's only reference to that dependency.

Verification

Built dist/ and ran unitTests/resources/dropTableCrossWorkerWrite.test.js under node_modules/.bin/mocha 100 consecutive times on main (50 before this edit, 50 after): 0 failures, and the second test's raced > 0 assertion — confirming the race was actually exercised (a record lock held by the worker's in-flight commit at the moment dropTable() was called) — passed on every run.

Close recommendation

Both #1381 and #2496 can close:

• Cross-worker write can race RocksDB table drop and poison catalog cleanup #1381 (cross-worker write races table drop, poisons catalog cleanup): the reproducer exercises exactly this shape — a worker's source-fill commit racing dropTable() on the main thread — across 20 iterations per run, 100 runs total, with the record lock sampled as held at drop time on every run and no "Invalid column family" / unexpected worker events / catalog residue in any of them.
• main is red: Integration Tests 3/6 fails on drop_table with 'Invalid column family specified in write batch' since #2492 #2496 (main red after Stop the RocksDB WriteBufferManager from stalling every writer by default #2492 removed the RocksDB write-buffer stall that had been masking this race): the underlying fix (rocksdb-js#850 deferred reclamation + harper's generation-distinct table stores) addresses the race Stop the RocksDB WriteBufferManager from stalling every writer by default #2492 exposed, not Stop the RocksDB WriteBufferManager from stalling every writer by default #2492 itself; no CI sighting of the signature since before the Defer column-family reclamation behind generation-distinct table stores #2604-era work and the 2.10.0 pin landed.

For the human reviewer

• Pre-push review (codex + gemini + cursor-composer + harper-domain, 3 rounds) landed on verdict: LGTM, no open in-scope findings.
• One outside-this-diff finding was carried through adjudication and deliberately left unaddressed here (out of this task's scope): unitTests/resources/dropTableGeneration.test.js:247 has the same columnFamily.pendingReclaims precondition, now the only remaining deterministic guard on the binding capability now that this PR's reproducer checks it only behaviorally (raced > 0). The adjudicator's read: that's fine as-is — it's an assert.ok, not a skip, and production (resources/Table.ts:422, resources/Table.ts:446, resources/databases.ts:5375) reads the same stat name, so main would already be failing if the name drifted — but whether to reword its "bump the pin" message now that this file no longer carries a matching one is a judgment call left to the reviewer.

Framing-Verdict: chosen-approach-sound

🤖 Generated with Claude Code

https://claude.ai/code/session_01BqdNKW8ja6cPc238RVJJYZ

_{Review-Coverage: authored=claude; ran=gemini,cursor-composer,codex; adjudicated=domain; declined=cursor-grok,cursor-kimi,cursor-muse; rounds=3; full=1 @ f9268f7}

_{Human-Review-Need: 3 (decisions: drop-binding-capability-precondition) @ f9268f7}

[gemini-code-assist]

Code Review

This pull request simplifies the unit test in dropTableCrossWorkerWrite.test.js by removing the unused database import and removing an assertion that checked for columnFamily.pendingReclaims in the database stats. A comment has been added explaining that the test now relies on @harperfast/rocksdb-js to defer physical column-family drops behind admitted commits. There are no review comments, so I have no feedback to provide.