Performance baselines

uvscope ships a dependency-free reference benchmark for the reverse proxy. The goal is a repeatable relative measurement on one machine — not a cross-machine guarantee. Absolute numbers depend on the CPU, kernel, Python client and the origin server.

What it measures

scripts/bench.sh starts:

1. a local HTTP/1.1 origin (Python ThreadingHTTPServer, Nagle disabled) on --origin-port;
2. a throwaway uvscope daemon on --port with a temporary data directory;
3. a reverse proxy on --proxy-port targeting the origin (in-process by default, or a child process with --isolated).

It then sends --requests sequential keep-alive GET requests to the origin directly and through the proxy, and reports throughput plus mean/p50/p95/max latency. Both paths use the same Python http.client connection and payload, so the difference is the proxy path itself.

./scripts/bench.sh ./build/uvscope --requests 500
./scripts/bench.sh ./build/uvscope --requests 500 --isolated
./scripts/bench.sh ./build/uvscope --requests 500 --json

The client is single-threaded: this is a latency/overhead measurement, not a saturation test. For concurrency or saturation numbers, drive the listener with your own load generator and compare release builds on the same host.

Reference run

Recorded on 2026-10-25 with the v1.0.0 release build:

• CPU / host: Linux 6.14.0-37-generic x86_64 (glibc 2.39), development container
• Proxy: in-process reverse proxy, empty rule set
• Requests: 500, warmup 50, keep-alive, sequential

Path	req/s	mean ms	p50 ms	p95 ms	max ms
Direct origin	6209	0.16	0.15	0.23	0.34
Reverse proxy (in-process)	567	1.76	1.76	1.88	2.59
Reverse proxy (isolated child)	577	1.73	1.73	1.84	2.88

Interpretation:

• The proxy adds about 1.6 ms per sequential request on this host.
• In-process and isolated proxies have effectively the same data path cost; isolation adds a child process and IPC telemetry, not per-request IPC. Choose isolation for fault containment, not for throughput.
• The direct-origin number is dominated by the Python client and loopback; use the ratio, not the absolute rate.

Record the command, host and commit alongside any baseline you publish. Numbers from different hosts or Python versions are not comparable.

Release verification

The release pipeline runs, per platform:

• --self-check (IPC, crypto, SQLite, CA, HTTP server, proxy-child);
• the full smoke suite on Linux;
• a Windows ConPTY terminal smoke and the H2 runtime check;
• a macOS self-check and the H2 runtime check.

The benchmark is a developer/operator tool; it is not a pass/fail CI gate because CI hosts are shared and noisy.