Benchmarks

The numbers below are the current output of the repository's own Go
benchmarks (client/bench_test.go and
server/internal/ingest/bench_test.go), run on a single developer machine
(Apple M3 Pro, go1.26.5, darwin/arm64). They characterize relative cost
and allocation behaviour, not a guaranteed production figure: absolute
numbers depend on hardware, Go version, and load. Reproduce them on your own
target hardware before relying on them for capacity planning.

What is measured

• BenchmarkObserve / BenchmarkObserveParallel (client/bench_test.go):
  the cost of Recorder.Observe on the steady-state path, where the series
  already exists, so only shard-select, lock, counter increments, and
  sketch.Add run. This is the cost added to every recorded request.
• BenchmarkUploadSingleSummary / BenchmarkUploadBatch /
  BenchmarkUploadParallel (server/internal/ingest/bench_test.go): the
  cost of the ingest handler's Upload RPC, from authentication through
  timestamp policy, cardinality/rate checks, and row conversion, for a batch
  of 1 and a batch of 64 summaries.

Measured results

Client, per-request hot path:

The steady-state path is allocation-free, matching the design goal described
in Runtime overhead: only the first observation of
a new series in a window allocates.

Server, ingest Upload RPC (per call, not per summary):

A client flushes once per flush window (10 seconds by default) regardless of
request volume, so this per-upload cost is incurred once per window per
instance, not once per request.

Reproducing these numbers

cd client && go test -run '^$' -bench 'BenchmarkObserve$|BenchmarkObserveParallel$' -benchmem ./...
cd server && go test -run '^$' -bench 'BenchmarkUpload' -benchmem ./internal/ingest/...