perf(backend): avoid full-snapshot parsing on unchanged empty-tail polls

未关闭
#2,479 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
25/100
Issue 类型
重构
描述清晰度
基本清楚
活跃度
活跃

调研方向

Start in packages/squad-sdk/src/backend/local-backend.ts: the readEvents filtering and slicing around lines 552-569, and the snapshot loading in load() around lines 878-903. Reproduce the empty-tail poll (readEvents with after set to the current tail and limit 1) against a built copy, and count how often the full snapshot is deserialized. Done means an unchanged poll performs zero full deserializations, with the external-append, restart and corrupted-persistence cases still passing and the benchmark evidence the issue asks for.

由索引模型根据 Issue 内容生成。

描述

Summary

The local file backend reads/parses the complete conversation snapshot before filtering/slicing readEvents results. A one-event request and an empty-tail poll therefore scale with the full conversation history rather than the requested page.

Proposed primary owner: sdk-runtime-dev.

Baseline evidence

Revision: 47082a81799710c3e457b594123d190c00596ebf, observed 2026-11-12.

Code: local-backend.ts:552-569, and snapshot loading at lines 878-903.

Independent source-function repeat:

History events p50, one-event tail page p50, empty-tail poll
100 0.110 ms 0.110 ms
1,000 0.849 ms 0.849 ms
10,000 9.353 ms 9.203 ms

Fixture: production local file-state backend, synthetic valid event/cursor sequence, after cursor and limit: 1, returning either the final event or no events. At 10,000 events the complete snapshot is 5,795,874 bytes. Assertions verified count, final sequence and resumability. A separate diagnostic CPU profile identified load() as the dominant sampled path; profiling was outside these timing runs.

Environment: Apple M4 arm64, 16 GiB, Darwin 25.6, Node 24.15.0; three rounds of ten sequential observations per case, two warmups, warm filesystem caches, setup/imports outside timing, rotated/reversed case order. Unmodified TypeScript methods ran through Node's built-in transformation and isolated relative-import resolution.

Limit: locked restoration failed with a configured-feed E404, blocking packaged build/tests. These are source-function diagnostics without a real harness/model, not application-throughput or optimization gains. Reproduce through the built path and publish portable fixtures/raw evidence before accepting a change.

Scope

Parent: github/squad#2481.

One hot path: repeated unchanged readEvents(after=current-tail, limit=1) calls against the local file-state backend. A one-event page is a correctness/control case, not a second optimization requirement. No public cursor redesign, harness/model changes, new storage backend or broad event-store rewrite.

Measurable acceptance and tests

For an optimization, use at least three interleaved built-baseline/change rounds with at least 30 observations per case per round; publish raw observations, per-round p50/p95, exact SHA/environment and correctly labeled CPU/I/O/memory evidence. The budgets below are proposed acceptance targets, not already-demonstrated gains. If a safe material win cannot be reproduced, record rejection/defer rather than broaden the implementation to chase the target.

  • Fixed histories: 100/1,000/10,000 valid events. Measure repeated empty-tail polls and a single newly appended event using the same cursor contract.
  • After initial hydration, an unchanged empty-tail poll must perform zero full conversation-snapshot deserializations. Instrument and assert that count separately from timing.
  • Target at least 25% lower p50 at 10,000 events than the same built baseline; publish p95, bytes read/parsed, CPU and retained heap at every size.
  • Append an event from another process, then verify the next poll observes it with the exact sequence/cursor/resumability contract. Include stream/limit controls and restart replay.
  • Test corrupted persistence, epoch/fencing change and cancellation/approval boundaries affected by the implementation. Preserve explicit errors and durable recovery; no stale or successful partial state.
  • If caching/indexing is used, assert bounded retained state and invalidation on external change/restart. Initial/recovery loads may parse fully but must be labeled separately from steady-state polls.
  • Publish a portable built-path fixture and before/after results; keep implementation limited to this polling path and necessary correctness plumbing.

Related work

The backend/harness implementation in github/squad#2451 is related prior art, not evidence of bounded polling performance. A fresh open-issue search found no focused readEvents performance duplicate.

Methodology: dotnet/aspnetcore#70219 and microsoft/waza#846.

主要语言
TypeScript
星标
3.4k
派生
516
平均合并
1 天 15 小时
30 天内合并 PR
116

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

github/squad 的其他 Issue

查看 github/squad 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。