Performance¶
Results below were measured on Apple M1 running macOS 26.6 and Go 1.26.6 on 2026-08-18. Run same commands on target hardware before capacity planning.
Zero-latency provider¶
scripts/loadtest_e2e.sh drives actual HTTP server, JSON handling, routing, input policy, usage tracking, and zero-latency mock provider.
Command: hey -n 30000 -c 50
| Metric | Result |
|---|---|
| Throughput | 54,860 requests per second |
| p50 | 0.6 ms |
| p95 | 2.4 ms |
| p99 | 3.5 ms |
| Errors | 0 of 30,000 |
Provider-latency test¶
scripts/benchmark.sh disables cache and sends every request to mock provider with fixed 25 ms delay.
Command: hey -n 300 -c 20
| Metric | Result |
|---|---|
| Throughput | 611.75 requests per second |
| p50 | 27.9 ms |
| p95 | 34.0 ms |
| p99 | 36.4 ms |
| Error rate | 0.00% |
Governance microbenchmarks¶
Go benchmark results:
| Scenario | Time per operation | Allocations |
|---|---|---|
| Envelope creation | 402 ns | 4 |
| Policy allow, 20 rules | 666 ns | 4 |
| Policy block, no match | 577 ns | 4 |
| Policy plus evidence | 2.634 us | 19 |
| Full allow with benchmark credential | 2.905 us | 22 |
| Review queue submission | 1.173 us | 5 |
Standalone percentile runner reported 0.67 us p50 for policy evaluation, 2.62 us p50 for policy plus evidence, and 2.58 us p50 for full allow path across 10,000 iterations:
Microbenchmark operations per second and HTTP throughput measure different paths. Provider latency, policy count, evidence backend, telemetry exporters, TLS, and concurrency affect deployment result.