Performance

Daravel makes no performance claim without a reproducible number (BLUEPRINT_V2 P10). The harness in bench/ builds the servers AOT and measures them:

cd bench
dart run bin/bench.dart --duration=10 --isolates=1,4 --out=results.json

It reports requests per second and p50/p95/p99 latency for:

1. GET /ping (framework overhead),
2. JSON plus one query,
3. validation plus an insert in a transaction,
4. eager loading two relations,
5. a throttled route (coordinator cost),
6. login (bcrypt in the compute pool),

plus a Zone/facade microbenchmark, startup time and idle RSS for each isolate count. Every scenario is also expressed relative to a plain dart:io server measured in the same run; the nightly workflow fails when that ratio drops by more than 10% against bench/baseline.json.

Internal targets (not public promises): /ping within 25% of dart:io, throttling adds at most 100 µs of service time per request, and four isolates give at least three times the throughput of one on a 4+ core machine. See bench/README.md for the methodology.