Benchmarks¶
Every public method is benchmarked two ways, with Go's testing.B
(bench_test.go). The numbers below were measured once, on the machine
named in Environment; they describe that machine on
that day, not a guarantee.
Method¶
BenchmarkClient/<Method>: the client's own cost, no network. A cannedhttp.RoundTripperanswers every request with a response body recorded from pain001 0.0.87 (testdata/bench/*.json, recorded byscripts/record-bench-responses.shwith the conformance kit's credit-transfer fixtures). Each iteration builds the request, JSON encodes the kit fixture into it (the 5-row CSV forValidateandGenerate, the 5-transaction pain.001.001.09 document forCheckandProjectFace), sends it through the transport, and reads and decodes the response into the method's Go type.BenchmarkLive/<Method>: one loopback round trip, the same calls againstpain001 serveon127.0.0.1. This is dominated by the server's work (parsing, validating, rendering XML) and includes the client's cost above. Skipped whenPAIN001_BASE_URLis unset.- Each benchmark ran 10 times for 1 second (
-count 10 -benchtime 1s), with-benchmem; the tables give benchstat's median and the spread of the 10 samples (±, a 95% confidence interval).
Reproduce:
pain001 serve --host 127.0.0.1 --port 8000 &
PAIN001_BASE_URL=http://127.0.0.1:8000 make bench BENCHCOUNT=10
go run golang.org/x/perf/cmd/benchstat@latest bin/bench.txt
CI smoke-runs every benchmark once per pull request
(make bench BENCHCOUNT=1 BENCHTIME=1x in make regression), to keep
them compiling and running; it asserts nothing about speed.
Environment¶
| Machine | Apple A18 Pro, 6 cores (GOMAXPROCS 6), 8 GiB, macOS 27.2 (26B123) |
| Go | 1.27.4, darwin/arm64 |
| Server | pain001 0.0.87 (ci/server-requirements.txt), Python 3.12.14, pain001 serve --host 127.0.0.1, no API key |
| Date | 2026-11-14, commit of this release branch |
| Load | Heavily shared: the 1-minute load average was 83 at the start and 50 at the end of the run, from other builds on the same machine. The wide ± spreads below come from that, not from the client. |
Results¶
Client-side cost (no network)¶
| Method | Time per call | Memory per call | Allocations |
|---|---|---|---|
Health |
5.7 µs ± 55% | 2.97 KiB | 26 |
Validate |
15.5 µs ± 186% | 5.02 KiB | 33 |
Generate |
104.8 µs ± 21% | 24.22 KiB | 43 |
Check |
36.0 µs ± 23% | 9.72 KiB | 35 |
Faces |
9.2 µs ± 19% | 3.39 KiB | 31 |
ProjectFace |
49.3 µs ± 25% | 26.91 KiB | 96 |
Request (GET health) |
6.5 µs ± 30% | 4.12 KiB | 44 |
Loopback round trip (pain001 serve on 127.0.0.1)¶
| Method | Time per call | Memory per call | Allocations |
|---|---|---|---|
Health |
0.78 ms ± 87% | 5.59 KiB | 67 |
Validate |
6.0 ms ± 92% | 8.01 KiB | 79 |
Generate |
16.2 ms ± 43% | 27.23 KiB | 89 |
Check |
10.6 ms ± 72% | 15.52 KiB | 83 |
Faces |
1.7 ms ± 70% | 6.00 KiB | 72 |
ProjectFace |
16.7 ms ± 43% | 34.09 KiB | 145 |
Request (GET health) |
1.0 ms ± 91% | 6.73 KiB | 85 |
Memory and allocation counts are stable across samples (spread 0 to 6%); times are not, on this loaded machine.
Reading the numbers¶
- The client's own work is microseconds per call; a round trip is
milliseconds, so for every method the client is under 1% of the call.
GenerateandProjectFacecost the most client-side because they decode the largest responses (the 3.5 KB XML document; the face payload and loss report). - The round-trip numbers are the server's, on loopback, with no TLS and
no network latency. They say nothing about a deployment; measure
yours with
BenchmarkLiveagainst it. - No number here is compared with another library: the suite has no other Go client for this API to compare with.