Performance

The performance target is predictable low memory rather than maximum request throughput. The gateway bounds request bodies and concurrent work and streams responses without retaining the full generation.

Always benchmark the exact provider, concurrency, request size, and container limit used in deployment.

For local stream and goroutine investigations, enable the separate diagnostics listener with GOFEATHERROUTE_PPROF_ADDR=127.0.0.1:6060. Capture profiles outside the repository and compare repeated runs with benchstat; do not treat one profile or one workload as a general performance claim. Check the goroutineleak profile after cancellation and idle-stream workloads.

The benchmark harness also records CPU, memory, network I/O, block I/O, process count, swap, and OOM/restart state for Go Feather Route and the pinned LiteLLM comparison image.