Performance Auditor

Overview

Find where time and memory actually go, prove it with measurements, and propose the highest-leverage optimizations. Rule number one: measure before recommending - no cargo-cult optimization.

Hot spots by layer

Layer	Typical issues
Database	N+1 queries, missing indexes, SELECT *, unbounded result sets, no connection pooling
Backend	blocking I/O inside async code, sync file/network calls in hot paths, quadratic loops, chatty logging
Caching	recomputing pure results, missing HTTP cache headers, cache stampedes
Frontend	oversized bundles, unoptimized images, render waterfalls, missing memoization, layout thrashing
Memory	leaks (listeners, closures, caches without eviction), large object retention
Startup	eager imports, synchronous config fetches, unbounded migrations

Workflow

1. Establish the baseline: what is slow, by how much, and for whom? Get a number first (timer, profiler, EXPLAIN ANALYZE, Lighthouse, time).
2. Profile with what's available:
   • Python: cProfile, py-spy, tracemalloc
   • Node: --cpu-prof, clinic, Chrome DevTools
   • SQL: EXPLAIN (ANALYZE, BUFFERS), slow query log
   • Web: Lighthouse, bundle analyzers, Web Vitals (LCP, INP, CLS)
3. Attribute cost: rank the top offenders by measured share of time/memory.
4. Propose optimizations sorted by impact / effort ratio, each with:
   • the measurement proving the problem,
   • the change,
   • the expected gain (estimate honestly),
   • the risk.
5. If asked to apply fixes: change one thing at a time and re-measure after each.

Anti-patterns

• Optimizing without a baseline measurement
• Micro-optimizations while an N+1 query dominates
• Adding caches without an invalidation story
• Claiming precise speedups you did not measure