Skip to content

performance-checks.md

Latest commit

 

History

History
38 lines (25 loc) · 2.3 KB

performance-checks.md

File metadata and controls

38 lines (25 loc) · 2.3 KB
Raw

Performance checks

Local benchmark smoke

make bench

This runs packages under internal/testkit, internal/core/stream, internal/core/runtime, internal/core/routing, internal/core/diag, and streaming encoders. For a single package, e.g. go test -bench=. -benchmem -run=Benchmark ./internal/core/runtime/.... Secure-session recorder smoke: go test -bench=BenchmarkRecorder -benchmem -run=^$ ./internal/core/securesession/app (also included in make bench).

Comparing before/after (benchstat)

go test -bench=BenchmarkExecutorExecuteAndDrain32Deltas -benchmem -count=6 -run=^$ ./internal/core/runtime/... | tee /tmp/bench-old.txt
# change code
go test -bench=BenchmarkExecutorExecuteAndDrain32Deltas -benchmem -count=6 -run=^$ ./internal/core/runtime/... | tee /tmp/bench-new.txt
benchstat /tmp/bench-old.txt /tmp/bench-new.txt

Profiling the streaming path (alloc space)

Example: allocation profile from the runtime benchmark (replace the mem profile path as needed):

go test -memprofile=mem.prof -bench=BenchmarkExecutorExecuteAndDrain32Deltas -benchtime=500ms -run=^$ ./internal/core/runtime/...
go tool pprof -top -nodecount=30 mem.prof

retryRecvStream.Recv uses diag.EnsureCallDiag, which allocates a child context only when the trace / A-leg on ctx differs from the active attempt. Bundled HTTP frontends call diag.EnsureCallDiag(ctx, call.ID, call.Session.ALegID) immediately after a successful Executor.Execute (prepareSubmitAndALeg sets call.Session.ALegID). Custom callers should do the same before draining EventStream from the HTTP request context if they want this fast path.

Baseline profiling note (executor benchmark)

Heap profiles of BenchmarkExecutorExecuteAndDrain32Deltas showed diag.WithCallDiag and context.WithValue among top allocators when Recv wrapped an empty request context each time. Combining the frontend EnsureCallDiag attach with resolver-populated call.Session.ALegID, the same benchmark drops to ~69 allocs/op for 32 deltas (was ~145 without the attach).

CI

The GitHub Actions workflow at .github/workflows/benchmarks.yml runs make bench on a schedule and on workflow_dispatch, and stores the text log as a build artifact (for optional benchstat against a saved baseline on a developer machine).