LMDB MDB_NOTFOUND freelist error causing uncaught commit exception under sharded write load

[kriszyp]

Problem

Under sharding write stress (75k req/s across 5-node cluster), LMDB throws MDB_NOTFOUND when attempting to delete a free-space record, which propagates as an uncaughtException crashing the process.

uncaughtException Error: Commit failed (see commitError for details)
    at rejectCommit (lmdb/write.js:612:17)
    at resolveWrites (lmdb/write.js:574:40)
    ...
Error: MDB_NOTFOUND: No matching key/data pair found: Attempting to delete free-space record
    at Function.<anonymous> (lmdb/write.js:516:7) {
  code: -30798
}

Relation

Closely related to #685 (LMDB freelist SIGSEGV under higher load). Both originate from LMDB freelist corruption under concurrent sharded writes — this ticket surfaces as an application-level uncaughtException; #685 surfaces as a native SIGSEGV.

Reproduction

Performance test repo: https://github.com/HarperDB/sharded-blob-test

Triggered by multiRightWrites.js (1 KB records, 5-node sharding, 4.6.0-alpha.3, 75k req/s total).

Impact

Unhandled commit exception terminates the server. The commitError is a pending Promise at the time of rejection, suggesting the LMDB write path doesn't cleanly surface freelist errors to the caller before crashing.

Next steps

• Investigate alongside LMDB SIGSEGV on freelist corruption under high-throughput sharded writes #685 — root cause likely the same freelist state corruption
• Determine if lmdb's commit error handling can catch and surface MDB_NOTFOUND as a recoverable error rather than crashing
• Evaluate back-pressure / write batching for high-concurrency sharding paths

🤖 Filed by Claude on behalf of Kris.