perf(api): add compact calculator benchmarks / perf(api)：新增计算器精简基准测试响应

[Oseltamivir]

Summary

• Add a cached, page-owned view=calculator&sequence=... response to /api/v1/benchmarks.
• Return only the selected calculator scenario and metric keys used by interpolation; omit per-worker telemetry from this view.
• Keep the default raw benchmark API response unchanged for inference charts and external consumers.
• Give each compact sequence its own React Query and Blob cache identity.
• Document the new backend-for-frontend exception and add projection, API-route, client URL, cache-key, and calculator regression coverage.

Impact

Against the repository's representative benchmark fixture, the uncompressed response is smaller by:

• 1k/1k: 78.0%
• 1k/8k: 88.2%
• 8k/1k: 78.0%

This reduces calculator transfer size, browser JSON parsing, and retained React Query memory without changing the reusable raw API contract.

Validation

• bun run lint
• bun run fmt
• bun run typecheck
• Focused unit/API/calculator tests: 120 passed
• Calculator Cypress E2E: 66 passed
• Full unit suite: 3,236 passed; one unrelated pre-existing timezone-sensitive failure remains in visit-tracking.test.ts (does NOT increment when a new session happens on the same calendar day).

中文说明

• 为 /api/v1/benchmarks 新增带缓存的页面专用响应：view=calculator&sequence=...。
• 仅返回计算器当前所选场景和插值所需指标，并在该视图中省略逐 worker 遥测数据。
• 推理图表和外部消费者使用的默认原始基准测试 API 响应保持不变。
• 不同精简场景分别使用独立的 React Query 与 Blob 缓存标识。
• 补充后端为前端服务（BFF）例外说明，以及字段投影、API 路由、客户端 URL、缓存键和计算器回归测试。

效果

基于仓库中的代表性基准测试 fixture，未压缩响应体缩小：

• 1k/1k：78.0%
• 1k/8k：88.2%
• 8k/1k：78.0%

该优化可减少计算器的数据传输量、浏览器 JSON 解析开销及 React Query 内存占用，同时不改变可复用的原始 API 契约。

验证

• bun run lint
• bun run fmt
• bun run typecheck
• 聚焦单元测试、API 测试及计算器测试：120 项通过
• 计算器 Cypress E2E：66 项通过
• 完整单元测试：3,236 项通过；visit-tracking.test.ts 中仍有一项与本改动无关、受时区影响的既有失败（does NOT increment when a new session happens on the same calendar day）。

---

Note

Low Risk
Scoped BFF addition with a separate cache namespace; default benchmarks API and inference chart fetching are untouched, with regression tests on projection, routing, URLs, and cache keys.

Overview
Introduces a page-owned compact response on /api/v1/benchmarks via view=calculator&sequence=..., documented alongside other deliberate BFF exceptions. Valid sequences are 1k/1k, 1k/8k, 8k/1k, and agentic-traces; unknown sequences return 400 before hitting the DB.

The route applies toCalculatorBenchmarkRows: filter to the requested scenario, drop workers, and keep only metric keys the calculator interpolation path uses. Results are cached under a separate benchmarks-calculator blob cache key; the default benchmarks response and inference chart consumers are unchanged.

Client plumbing extends fetchBenchmarks, React Query benchmarkQueryOptions / useBenchmarks (sequence-specific cache keys), and useThroughputData so the calculator requests the compact view per selected sequence instead of full raw rows.

^{Reviewed by Cursor Bugbot for commit ff367db. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 10, 2026 7:40am