LMDB MDB_BAD_VALSIZE / mdb_page_alloc error under sharded blob write load

[kriszyp]

Problem

During blob write stress tests on a sharded cluster, LMDB fails with MDB_BAD_VALSIZE and related page allocation errors, ultimately causing an uncaught commit exception.

Log sequence:

Read transaction detected that has been open too long (over 15 minutes), ending transaction
Invalid zero size key
mdb_page_alloc error
mdb_page_touch error
cursor_del0 failed, -30781

Error: MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size: Attempting to delete free-space record
  code: -30781

Relation

Part of a cluster of LMDB freelist failures under sharding write load:

• LMDB SIGSEGV on freelist corruption under high-throughput sharded writes #685 — SIGSEGV from freelist overlap
• LMDB MDB_NOTFOUND freelist error causing uncaught commit exception under sharded write load #686 — MDB_NOTFOUND on freelist delete

All three share the same reproduction path (sharded-blob-test, multiRightWrites.js, multi-node cluster). The stale read transaction noted in logs likely contributes by preventing freelist page recycling, leading to state corruption.

Reproduction

https://github.com/HarperDB/sharded-blob-test — multiRightWrites.js and singleBlobWrites.js. Reproduced on both 4.5.9 and 4.6.0-alpha.3.

🤖 Filed by Claude on behalf of Kris.

[kriszyp]

Appears to be a duplicate of / consolidate under #686 — same LMDB freelist-corruption cluster (MDB_BAD_VALSIZE vs MDB_NOTFOUND).

— Claude (Opus 4.8), issue-backlog triage