gap CLI: DB's ~90 s static analysis is silent past the daemon wait, so every DB daemon check falls back cold

未关闭
#173 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
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
领域
cli, performance

调研方向

Start with check.mjs, analysis.mjs, and daemon.mjs to trace how analysis blocks the daemon heartbeat and triggers the cold fallback; compare the two proposed approaches. Build the issue’s required oracle-first end-to-end history using a test-only pause point to hold analysis past the CLI wait. Done when the daemon check is served with unchanged reuse and no “runs cold” note, and the witness fails on main.

由索引模型根据 Issue 内容生成。

描述

Summary

On TanStack DB's packages/db, the gap CLI's static analysis phase takes about 90 s, and the daemon sends nothing during it. #131 made the CLI check cold when a daemon has sent nothing for 60 s (WAIT_MS, daemon.mjs). So on DB the daemon never serves a check: the CLI always falls back cold (v21.2 R39), and the daemon default (R13) saves nothing there.

Measurement

Arca pool, Linux, 8 vCPU. DB at 21c52a07, --package packages/db --source src/query/subset-dedupe.ts. Phase durations come from the CLI's progress lines (pool tag do-dbphase). Each row is one run.

CLI warm contexts find test files static analysis
main 1.1 s 17 ms 91.2 s
main 0.7 s 21 ms 88.9 s
#133 head 0.8 s 17 ms 91.4 s
#133 head 0.8 s 17 ms 91.6 s
  • The time is the same on main and on #133, so #133 did not cause it.
  • The DB daemon bench for #133 had to raise the CLI's wait with the test-only knob ORACLE_GAPS_TEST_WAIT_MS for the daemon to serve any check.

Cause

  • checkIn (check.mjs) calls analyze (analysis.mjs) synchronously on every check. analyze instruments the sources and then calls ts.createProgram over the sources and the files that owns names. On DB, that means the type checker reads a large part of the package.
  • While it runs, the event loop is blocked, so the daemon cannot send its beat.
  • daemon.mjs's send already names this case: "a phase with no running copy (a session's start, a warm-up, the analysis) that runs longer than the bound".
  • The analysis runs in each check, not once per session. So this probably affects every daemon check on DB, not only the first. This is not measured separately.

Proposed fix (choose one)

  1. A beat around the analysis. The daemon sends a beat at each phase boundary, and also from inside the analysis between its steps: after each source is instrumented, and before and after createProgram. This is small. But createProgram alone can still run past 60 s on a big package, so it does not close the case by itself.
  2. Analysis off the main thread. Run analyze in a worker thread, so the daemon's event loop stays free to send beats on its normal schedule. This closes the case for any package size. It also lets a later change keep the ts.Program between checks (oldProgram), which would make rechecks cheaper.

Either fix needs a witness that is oracle-first (P9): a constructed e2e history whose analysis phase is held past the CLI's wait at a test-only pause point (R57, never by time), and whose daemon check must still be served (unchanged reuse, no runs cold note). It must fail on main.

Authority

  • R13 / M-30: the daemon is the default.
  • R39: a cold fallback is not a failure, but it must stay the exception.
  • Raised in the review of #131 and measured during #133.

This issue was written by Isaac.

主要语言
JavaScript
星标
5
派生
1
平均合并
11 小时 18 分钟
30 天内合并 PR
66

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

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。

TanStack/oracles 的其他 Issue

查看 TanStack/oracles 的全部 Issue

相似的 Issue

更多 JavaScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。