Freegle / Iznik / 24904 / 24904

72%

master: 72%

Pull #958

circleci

perf(markseen): batch the "mark seen" inserts (Galera-safe) instead of one INSERT per post

"Mark seen" on Browse ran one INSERT ... ON DUPLICATE KEY UPDATE per message id, so it scaled
linearly with the unread count - one DB round-trip per post. Batch the inserts into multi-row
statements (chunk 100): a mark-seen of N posts is now ~ceil(N/100) round-trips. Per-message
semantics are unchanged (ON DUPLICATE KEY UPDATE is per conflicting row).

Galera (Percona XtraDB Cluster) safety - a multi-row INSERT is one transaction, so a
certification conflict rolls back the whole statement (unlike the old per-row loop):
- Moderate chunk (100) bounds the rollback blast radius and lock footprint.
- Sort + de-dupe the ids so concurrent same-user mark-seens lock rows in a consistent order
  (no deadlock from opposite orderings) and a repeated id can't collide with itself in one
  statement.
- database.RetryExec retries transient connection/WSREP errors.
- Per-row fallback on deadlock/lock-timeout: a conflicting chunk degrades to per-row inserts
  (each its own tiny autocommit txn, the old behaviour) so the rest still land.

Follow-up (separate, larger): make mark-seen O(1) via a per-user, per-view high-water mark and
move the per-item View writes (needed for view-counts, dedup, the rippling underexposure score)
to the background_tasks queue.

Go suite green (3385/0): new TestMessagesMarkSeenBatchPreservesSemantics + TestDedupeSortedIDs;
existing markseen tests unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014ZeG2zmg2JCR8pYEjAUaAk

574 of 2797 branches covered (20.52%)

Branch coverage included in aggregate %.

355 of 1288 relevant lines covered (27.56%)

13.19 hits per line