rails-performance

# Rails Elixir Performance Establish functional parity for the measured operation first. Freeze before/after revisions or images and use the same fixture data, actual user state, release settings and dependency services. Benchmark application responses, not login redirects, error pages or an empty path presented as a populated workload. Record expected statuses, transport/HTTP errors and workload. For deterministic process/build-script comparisons use the toolkit benchmark command with balanced even rounds and --compare-output when the output should match. It records process wall time, not HTTP throughput. For servers use a suitable load generator with controlled clients, warmup, alternating order, CPU placement and raw per-round measurements. Measure throughput, latency percentiles, CPU per successful response, idle/peak memory, startup readiness and image size where relevant. State process RSS/PSS versus container/cgroup and socket/kernel memory, compressed versus unpacked images, and what is excluded. For DB work record warmed query counts as well as timing. Avoid compiling or running another load test during timed rounds. Repeat only when a changed workload or unresolved concern warrants it. Find the actual hot path before simplifying libraries, pools, caches or allocation. Prefer existing direct queries, cached statements, shared compatibility primitives and bounded buffers. Measure BEAM scheduler, garbage collection, native parser and compression costs before changing runtime settings. Do not port SQLite-specific WAL/pool wins to MySQL by analogy. Read [measurement.md](references/measurement.md) for workload/capacity traps and the measured optimizations that transferred between these ports. Keep raw before/after evidence under bench/results and update tables with the current state, with micro and end-to-end gains separate. Forecasts need labeled assumptions and a real target-hardware test; they are not benchmark wins.