The previous deploy of c4f6cc1 failed because I had applied the timings
column manually via psql before the deploy ran, leaving the hash-based
migrate runner with raw ADD COLUMN that conflicts with the existing column.
Hash differs after this change so the runner will retry on next deploy
and succeed (no-op on the already-applied column, success path records
the hash).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Production p99 was 18min for failed decodes and 15min for PL24 successes;
same VIN could trigger N parallel 10+ min decodes. Three layered fixes:
- In-flight dedup via Redis SETNX (vin🔒*); concurrent same-VIN requests
poll the resolve cache instead of re-firing the upstream chain.
- 24h positive cache (vin:resolve:*) and 6h negative cache
(vin:resolve:neg:*); previously 5min positive / no negative.
- 25s hard abort budget via AbortController; PCAT gets the signal natively
(AbortSignal.any), PL24/EMEX wrapped with raceWithSignal at the boundary.
Aborted decodes don't poison the negative cache.
- PCAT/EMEX real race: first definitive single-result wins; the slower
source is skipped (previously PCAT was always awaited first).
- PL24 circuit breaker: 3 consecutive failures opens a 30s cooldown
(pl24:cb:cooldown_until); successes reset the counter.
- Stage-level timings in query_logs.timings (jsonb): pcat/emex/pl24/
lock_wait/cache_hit/aborted. Failed source now logged as "none" or
"aborted" instead of misleading "corgi".
Verified locally with 3 parallel decodes of a fresh VIN: 1 real decode
(3.59s), 2 lock-waits (3.53s) sharing the result, 4th request 23ms cache
hit. Previously this would have been 3 separate 10+ min PL24 decodes.
Migration 0002 adds query_logs.timings jsonb (NULL default). Must be
applied manually before deploy (deploy.sh does not run db:push).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
