28 lines
2.4 KiB
Markdown
28 lines
2.4 KiB
Markdown
# Local backend A/B benchmark
|
|
|
|
Measured 2026-10-14 on the same Linux host, Rust optimized release build vs Bun/Hono, loopback HTTP, Chromium not involved in the load generator. Both services served the identical `frontend/` tree and used a seeded SQLite search row plus a 1 MiB cached-media fixture. The benchmark ran sequentially, four routes, 750 ms per route/concurrency. `responseBytes` are Node fetch's decoded response-body bytes, not raw compressed wire bytes. These are short smoke measurements; use a longer run and production-like media/network before drawing capacity conclusions.
|
|
|
|
Command:
|
|
|
|
```sh
|
|
BENCH_QUERY=rust-backend-parity-20261014 TEST_MEDIA_ID=dQw4w9WgXcQ node alt/bench.mjs http://127.0.0.1:3101 --duration-ms=750
|
|
BENCH_QUERY=rust-backend-parity-20261014 TEST_MEDIA_ID=dQw4w9WgXcQ node alt/bench.mjs http://127.0.0.1:3100 --duration-ms=750
|
|
```
|
|
|
|
Each cell is `requests/sec (p95 ms)`; all 32 route/concurrency runs reported zero errors.
|
|
|
|
| Concurrency | Backend | `/api/version` | CSS asset | Cached search | 1 MiB media range |
|
|
|---:|---|---:|---:|---:|---:|
|
|
| 1 | Bun | 2,425 (0.7) | 521 (3.4) | 1,041 (2.5) | 612 (3.5) |
|
|
| 1 | Rust release | 1,844 (0.9) | 665 (2.0) | 2,269 (0.6) | 480 (3.9) |
|
|
| 10 | Bun | 7,008 (2.7) | 1,179 (12.4) | 1,623 (17.2) | 685 (27.0) |
|
|
| 10 | Rust release | 3,872 (3.3) | 1,131 (12.9) | 4,228 (3.1) | 599 (27.4) |
|
|
| 50 | Bun | 10,033 (11.3) | 1,279 (44.3) | 1,783 (54.7) | 769 (105.5) |
|
|
| 50 | Rust release | 4,267 (17.4) | 1,199 (52.7) | 4,733 (14.5) | 689 (135.4) |
|
|
| 200 | Bun | 9,947 (33.7) | 1,333 (203.0) | 1,863 (161.0) | 1,029 (266.3) |
|
|
| 200 | Rust release | 4,396 (59.8) | 1,328 (202.7) | 4,533 (59.9) | 863 (644.7) |
|
|
|
|
The Rust backend led the cached-search route in this short run and was close on static CSS at high concurrency. Bun served `/api/version` and large media ranges faster at high concurrency; Rust's 1 MiB Range p95 widened to 644.7 ms at 200 connections. The shared in-memory static snapshot and boot-time Brotli/gzip encodings avoided disk lookup and per-request recompression, but the Range implementation still streams through Tokio file reads. No result implies production superiority.
|
|
|
|
Bun's `yt-dlp` Python worker could not import the system `yt_dlp` package on this host. Bench routes were seeded and do not exercise YouTube upstream latency or media encoding. Docker is unavailable on this host, so image build/compose smoke was not run.
|