decentraland / world-storage-service / 28621152435

99%

main: 99%

Pull #111

github

fix: close cache staleness under concurrent writes and a migration abort

Two issues found in a deeper review of the hardening changes:

- Cache coherence: the post-commit write-through set ran outside the quota
  advisory lock, and deleteValue takes no lock, so a concurrent delete's cache
  removal could land before a just-committed upsert's set — leaving a stale
  value served for the full TTL against a row that no longer exists. Replace
  the write-through with a post-commit invalidation (remove): removals are
  idempotent and order-independent, so the next read always repopulates from
  committed state. Renames the adapter hook cacheValue -> invalidateValue.

- Lowercase-world-names migration: the dedup only removed a mixed-case row when
  an already-lowercase twin existed, so two distinct mixed-case spellings of the
  same world (MyWorld / MYWORLD) sharing a (place_id, key) with no lowercase row
  collided on the UPDATE and aborted the migration. Deduplicate by keeping one
  row per lowercased key group (prefer an existing lowercase row, else the most
  recently updated), verified against Postgres for both cases.

Adds an integration test asserting a re-upsert through storage-operations is
visible to the next read (guards the post-commit invalidation).

351 of 362 branches covered (96.96%)

Branch coverage included in aggregate %.

170 of 174 new or added lines in 30 files covered. (97.7%)

1073 of 1079 relevant lines covered (99.44%)

71.47 hits per line