Benchmarking & Load Testing

Coordinated omission is when your load tool stops sending requests exactly when the server stalls, so closed-model wait-for-response tools hide tail latency while open-model constant-arrival-rate tools expose the true p99.

A trustworthy JVM benchmark warms up until C2 has compiled the hot path and GC is steady, pre-initialises pools and caches, records every latency into HdrHistogram, and isolates the one component under test.

A fair reactive-versus-blocking benchmark must defeat coordinated omission, avoid measuring a saturated shared database, match and correctly size connection pools, and size the young generation so R2DBC's tiny-object allocation does not create unfair GC pressure.

Chaos engineering is a controlled fire drill under real load: you break things on purpose, watch your SLIs, and prove the system heals itself.