mirror of
https://github.com/musistudio/claude-code-router.git
synced 2026-08-16 03:53:23 +00:00
30d1f35926
Some checks failed
CI / Typecheck (push) Has been cancelled
CI / Test architecture (push) Has been cancelled
CI / Test cli (push) Has been cancelled
CI / Test core (push) Has been cancelled
CI / Test electron (push) Has been cancelled
CI / Test ui (push) Has been cancelled
CI / CLI web E2E (push) Has been cancelled
CI / Docker system smoke (push) Has been cancelled
32 lines
1.5 KiB
Markdown
32 lines
1.5 KiB
Markdown
# Request log benchmarks
|
|
|
|
Run the request-log benchmark with the same Electron/Node native runtime selection used by the test runner:
|
|
|
|
```sh
|
|
npm run benchmark:request-logs -- --label local
|
|
```
|
|
|
|
To persist a machine-readable result:
|
|
|
|
```sh
|
|
npm run benchmark:request-logs -- --label local --output benchmarks/results/request-log-local.json
|
|
```
|
|
|
|
The benchmark has two layers:
|
|
|
|
- A request-log microbenchmark covering small records without a route trace, small records with a 10-hop route trace, 2 MB request bodies with a 10-hop trace, and 8 MB Base64 image requests.
|
|
- A real loopback HTTP service benchmark comparing no logging, synchronous main-thread SQLite logging, and the asynchronous worker logger.
|
|
|
|
The HTTP benchmark uses a separate load-generator process so client work does not share the server event loop. It reports QPS, durable QPS, p50/p95/p99/max latency, HTTP errors, rejected log events, event-loop p99/max delay, CPU, heap, and RSS changes.
|
|
|
|
The default web load is 10,000 requests with concurrency 64 and a 1 KB request body. It can be adjusted without editing code:
|
|
|
|
```sh
|
|
npm run benchmark:request-logs -- \
|
|
--web-requests 20000 \
|
|
--web-concurrency 128 \
|
|
--web-body-bytes 4096 \
|
|
--label web-load
|
|
```
|
|
|
|
Use `--skip-web` when only the storage microbenchmark is needed, or `--skip-storage` when iterating only on HTTP performance. Both layers report request-path time separately from durable flush time so moving work off the request thread cannot hide storage regressions.
|