cgw-performance

name

cgw-performance

description

Use when a change in safe-client-gateway adds or modifies a cached read, a cache key, a TTL, an outbound upstream call, a datasource, a BullMQ job, or anything on a hot path. Covers the new-cached-call recipe (CacheRouter getter, CacheFirstDataSource with a config TTL and notFoundExpireTimeSeconds, clear*() on every write, EventCacheHelper wiring - all in one PR), the never-call-fetch-directly rule, Promise.all/allSettled over sequential awaits, batching to provider limits, and the small-single-concern-measured protocol for perf PRs. Triggers on "cache", "Redis", "TTL", "invalidate", "slow", "p95", "latency", "N+1", "timeout", "retry", "backoff", "circuit breaker", "batch", "chunk", "upstream call", "queue job", "await in a loop".

CGW Caching and Performance

Read docs/agents/caching-and-performance.md before the change. This skill is a loader; the doc is the content.

Three non-negotiables from AGENTS.md live here:

• Never call fetch or an HTTP client directly. Outbound calls go through INetworkService / CacheFirstDataSource, or they silently lose logging, the circuit breaker, timeouts, and the error funnel.
• Cache keys only via CacheRouter; TTLs only from config. A hard-coded timeout, batch size, or TTL is a review finding.
• A cache write and its invalidation land in the same PR. The getter, the read, every clear*(), and the EventCacheHelper wiring are one change. Raising a TTL without invalidation is a deliberate, stated trade-off, not a default.

Perf PRs follow their own protocol: small, single-concern, and measured. A 19-file perf refactor here (#2926) was reverted the same day (#3034) because the one bad change could not be separated from the good ones.

Also load cgw-remarks — R-020 (bypassing NetworkService), R-021 (cache write with no invalidation), and R-022 (sequential or partial cache clears) are the recurring shapes.