perf: Squad performance audit rollup at 47082a8

オープン
#2,481 コメント 2 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
6/100
issue の種類
リファクタリング
明瞭さ
説明が足りない
活発さ
活発
技術スタック
nodejs, typescript

調査の方向性

This is a tracking rollup, not a task, and it names no source files or tests. Start from one of the linked focused issues (#2476 readEvents benchmark or #2482 startup baseline) and read its own scope and result contracts. Done here means every linked item has an evidence-backed disposition, per the close condition.

索引モデルが issue の本文から書いたものです。

説明

Summary and status

Track the measurement-first Squad performance effort at 47082a81799710c3e457b594123d190c00596ebf, observed 2026-11-12. This tracking-only rollup links seven focused investigations derived from the initial cross-package audit.

Status: baseline triage and issue creation only. No optimization implementation, shipped gain, or end-to-end speedup is claimed.

Proposed primary owner: architect; each focused issue names one accountable specialist.

The approach follows dotnet/aspnetcore#70219, Shayne Boyer's whole-product audit microsoft/waza#846, the real packaged-command correction in microsoft/aspire#16704, and total-session measurement criteria in microsoft/aspire#20873.

Focused findings and tracking checklist

  • github/squad#2476 — One built production readEvents scaling benchmark: six size/cursor cases, raw samples, assertions and calibrated regression-sensitivity tests. No framework/CI expansion. Owner: unit-contract-tester.
  • github/squad#2482 — Measurement-only built help/version startup baseline; three rounds of 20 fresh processes per command, elapsed/RSS data and runner failure tests. No startup optimization claimed. Owner: cli-dev.
  • github/squad#2477 — Governed local-memory search. At 5,000 active 1-KiB entries, source diagnostics measure p50 approximately 377-392 ms and 5,003 reads/query; preserve governance, matching and cross-process freshness. Owner: state-storage-dev.
  • github/squad#2475 — Latest-session discovery. At 1,000 retained archives, 1,001 synchronous reads and 75.2 MB read; independent daily-archive repeat p50 121.6 ms. Preserve eligibility, ordering and recovery. Owner: state-storage-dev.
  • github/squad#2479 — Unchanged empty-tail polling only. With 10,000 events, even an empty-tail poll parses a 5.8-MB snapshot and costs p50 approximately 9.2 ms. Preserve cursors, replay, fencing and lifecycle correctness. Owner: sdk-runtime-dev.
  • github/squad#2480 — Shell-history autosave. At 10,000 messages, each synchronous save rewrites 3.75 MB and takes p50 approximately 8.4 ms in the independent repeat. Preserve all history and crash/recovery behavior. Owner: state-storage-dev.
  • github/squad#2478 — Hosted CI test critical path. Three successful runs spent 301/328/394 seconds executing tests versus 8/9/13 seconds building. Profiling-only: same-SHA hosted profiles, attribution/top-ten table, unchanged inventory and at most one separately scoped follow-up. No sharding or optimization is part of this issue. Owner: release-engineer.

Measurement provenance and limitations

Runtime diagnostics used Apple M4 arm64, 16 GiB, Darwin 25.6, Node 24.15.0 / npm 11.12.1. Each timing case used three baseline rounds of ten sequential observations, two warmups, warm OS filesystem caches, and setup/imports outside timing. Cases were rotated/reversed between rounds. An independent coordinator pass reproduced 450 timing observations across memory, backend and session scenarios and verified output/count assertions, exact SHA and cleanup.

These are source-function diagnostics, not packaged startup results: missing TypeScript/Vitest and a configured-feed E404 for locked hono-4.13.17.tgz blocked dependency restoration, build and test execution. Missing dist also blocked built benchmark/cold-start runners. Node's built-in TypeScript transformation and an isolated relative-import hook were used for diagnostics; session-store imports were bound directly to its two actual production SDK exports rather than loading the full SDK barrel. This fallback does not satisfy release/optimization acceptance.

Timing, separate untimed I/O instrumentation, retained JS heap, allocation totals and total-session RSS are different measures. No paid/live inference was run. No allocation-total, token-efficiency, application/model-throughput or whole-agent-session memory improvement is established. Hosted CI observations are separate success-only runs on different commits/runners, not controlled A/B comparisons or reliability estimates.

Audit harness, raw samples, logs and a backend CPU profile are retained in the originating Copilot session. Portable checked-in reproduction fixtures and shared raw artifacts remain a requirement of the focused work; private local artifact paths are not a public reproduction dependency.

Scope boundaries

This issue is tracking-only, not an implementation epic to execute wholesale. Each linked issue has one bounded operation or diagnostic deliverable, explicit fixture sizes, result/test requirements and exclusions. Benchmark breadth was split into backend readEvents (#2476) and packaged help/version startup (#2482); other runtime fixtures stay with their owning optimization issues. CI #2478 ends at attribution, not an open-ended optimization sweep. Proposed percentage targets are experiment acceptance budgets, not established product SLOs or already-measured gains. If a target is unsafe or unsupported, record rejection/defer rather than expanding scope.

Method and compatibility gates

  • Check existing open issues, PRs and merged equivalents before advancing any new candidate. Each win should be a focused change, not a broad speculative refactor.
  • First restore/build the exact baseline and reproduce the affected real production path. Measure shipping/packaged commands separately from dev/source-import probes.
  • Before accepting each optimization, run at least three interleaved baseline/change rounds with equivalent workloads, raw observations and CPU plus correctly labeled allocation/retained-memory evidence. Require a material improvement beyond observed variation; record rejected/noise-level candidates.
  • Preserve public APIs, behavior, governance, history, isolation, freshness, explicit errors and recovery. No unsafe shortcuts, weakened tests, dropped messages or stale-cache success paths.
  • For any cache/pool/index, document bounded ownership/lifetime, invalidation and recovery across mutations, restarts and other processes. Account for required child processes before claiming total-session memory savings.
  • Run focused correctness/lifecycle/recovery tests and required checks on the actual candidate head. Separate evidenced baseline/flaky failures from introduced regressions.
  • Record exact SHA, environment, build mode, fixture sizes, cache/warmup policy, sample count and known coverage gaps with every result.
  • Keep implementation and merge decisions separate: no change is accepted from microbenchmarks alone, and no PR is merged without the repository's maintainer/review gates.

Deduplicated and deferred candidates

  • Resolver memoization, parallel charter discovery and asynchronous scheduler execution already landed in github/squad#1112, github/squad#1113 and github/squad#1114; do not recreate those issues.
  • Benchmark foundation github/squad#1111 is existing infrastructure, not missing work. The new benchmark issue is limited to readEvents evidence; startup measurement is separate. Model-quality paired A/B github/squad#2449 remains separate.
  • Generic context/tiered-memory proposals overlap github/squad#884, github/squad#1476 and github/squad#2392. The focused local-memory search issue should link to the ADR, not compete with it.
  • Resolver-cache stress retained approximately 11.8 MiB JS heap after 30,000 synthetic distinct paths and TTL expiry. Deferred, no focused issue created: real caller cardinality/materiality has not been established.
  • Eager CLI import optimization and SQLite amplification remain unmeasured hypotheses; no proven-bottleneck issue is proposed for them.

Close condition

Close only when every linked focused item has an explicit, evidence-backed disposition: accepted diagnostic/baseline deliverable with its specified result contracts and tests, accepted material improvement with real-path before/after evidence and correctness checks, or documented rejection/defer/duplicate with rationale. A green check, closed issue, moved allocation or lower synthetic benchmark alone is not completion.

主要言語
TypeScript
スター
3.4k
フォーク
516
平均マージ
1日 15時間
マージ済み PR(30日)
116

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

github/squad のほかの issue

github/squad の issue をすべて見る

似ている issue

TypeScript の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。