performance-optimization

N+1 queries are fixed by replacing per-record lookups with a single query using joins or ORM includes, such as findMany with an include clause. Confirm the fix by checking the database query log before and after the change.

Improve LCP by optimizing hero images with responsive srcset, fetchpriority, and correct dimensions; improve INP by reducing main-thread long tasks and unnecessary re-renders. Measure with the web-vitals library and Lighthouse, targeting LCP under 2.5s and INP under 200ms.

Add an index only after running EXPLAIN ANALYZE shows a sequential scan or stale statistics on the filtered columns. Index for the query shape with equality columns first, and skip indexing for low-selectivity filters, leading wildcards, or write-heavy tables.

Pool exhaustion shows as all endpoints slowing at once while the database sits idle. Raising the pool size just moves the queue; instead find what holds connections, and use a proxy like pgbouncer for serverless or autoscaling environments.

Do not optimize without profiling evidence of an actual bottleneck. Premature optimization adds complexity that costs more than it gains, and changes that measure as neutral should be reverted rather than kept.

Set explicit budgets for bundle size, API p95 latency, and Lighthouse scores, then enforce them with bundlesize and Lighthouse CI in the pipeline. Complement synthetic gates with real-user monitoring alerts on p75 metric movement.