claude-code-router/benchmarks/results/request-log-comparison.md

Request log performance comparison

Environment: Apple arm64, Electron Node v24.16.0, SQLite WAL, identical benchmark scenarios and batch sizes.

Scenario	Metric	Baseline	Optimized	Change
Small, no trace	Request-path throughput	9,949 records/s	173,695 records/s	17.5x
Small, no trace	Durable throughput	9,949 records/s	34,267 records/s	3.4x
Small, no trace	Event-loop p99	4.24 ms	1.61 ms	-62.1%
Small, 10 hops	Request-path throughput	5,922 records/s	66,716 records/s	11.3x
Small, 10 hops	Durable throughput	5,921 records/s	17,640 records/s	3.0x
Small, 10 hops	Event-loop p99	7.78 ms	1.30 ms	-83.3%
2 MB, 10 hops	Request-path throughput	126 records/s	553 records/s	4.4x
2 MB, 10 hops	Durable throughput	126 records/s	336 records/s	2.7x
2 MB, 10 hops	Event-loop max	204.08 ms	2.04 ms	-99.0%
2 MB, 10 hops	RSS delta	130.20 MB	66.58 MB	-48.9%

The optimized 2 MB result above was measured with the then-current 512 KB default, so it is intentionally marked degraded. Production now retains up to the 50 MB hard safety ceiling for each request/response body by default while preserving original size and truncation metadata.

Raw measurements:

• request-log-baseline.json
• request-log-optimized.json

HTTP service benchmark

The HTTP benchmark sends 10,000 POST requests with concurrency 64 and a 1 KB body from a separate load-generator process.

Mode	QPS	Durable QPS	p50	p95	p99	Event-loop p99	Errors	Rejected logs
No logging control	32,878	32,877	1.93 ms	2.34 ms	3.02 ms	1.96 ms	0	0
Synchronous main-thread SQLite	9,609	9,603	5.67 ms	11.44 ms	13.51 ms	10.53 ms	0	0
Asynchronous worker logger	31,879	25,757	2.02 ms	3.01 ms	6.20 ms	2.06 ms	0	0

Compared with synchronous main-thread logging, the asynchronous worker delivers 3.3x request QPS, 2.7x durable QPS, 54.1% lower HTTP p99 latency, and 80.5% lower event-loop p99 delay. Its request-path QPS is 3.0% below the no-logging control, and all 10,000 log events were accepted.

The first unrestricted-network run also exposed a linear scan through the downloaded model-pricing catalog, which limited durable logging to roughly 109 records/s. Building a source-and-model price index once per catalog load removed that bottleneck before the final measurements above.

Base64 image benchmark

The image scenario reuses an already-received 8 MB JSON request body containing an inline Base64 image, matching the point where request logging starts. Ten records were accepted and compacted. Request-path throughput was 922 records/s, durable throughput was 609 records/s, and no RSS growth was observed in the measured run. The stored JSON retains its structure, MIME information, original wire size, and an image-size descriptor without retaining the encoded pixels. MIME and encoding context are scoped to the same JSON object, so adjacent audio payloads are not compacted.

Raw measurement: request-log-base64-image.json.