First file search in a project waits on a second walk of every folder to find symlinks: 551 ms of a 1.25 s first @-mention on Windows

未关闭
#39,625 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。

  • #39644 来自 @juliusmarminge —— 已关闭,未合并

评估

难度
4/5
预计耗时
3-5 天
新手友好度
55/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
停滞

调研方向

Read apps/server/src/workspace/WorkspaceSearchIndex.ts, focusing on make, searchNames, and scanSymlinkEntries; the issue reports that symlink discovery rereads 4,703 indexed folders and blocks the first search. Start by tracing how fff builds its index and how symlink results are filtered after loadIgnoredPaths. Reproduce with the cold-search steps in the issue and compare the first-search trace; done means the search avoids or reduces the extra directory walk while still finding symlinks.

由索引模型根据 Issue 内容生成。

描述

What is slow

The first file search in a project (the first @ mention in the composer, or the first file-picker query) waits until the server has built that project's search index. Most of that wait is spent finding symlinks. The server reads every folder the index already holds a second time, with one readdir per folder, and the search does not answer until that walk is done. The index is dropped after 15 minutes without a search (WORKSPACE_INDEX_IDLE_TTL), so a user pays this again whenever they come back to a project.

Measured

  • Windows Server 2025, 4 vCPU, 16 GB, NTFS, Git for Windows, Node 26.13.1. Linux comparison: Linux VM, ext4.
  • Commit f2d4df0367.
  • Seed: a checkout of this repository itself, 32,715 tracked files in 4,702 folders, 17 tracked symlinks.
  • Five runs per platform. Each run starts a fresh server on a fresh data dir (T3CODE_DEV_SEED=0), then sends one projects.searchEntries with query manager, limit 100. Times are server span durations from server.trace.ndjson.
Windows median (range) Linux median (range)
ws.rpc.projects.searchEntries (first search) 1,253 ms (1,228-1,425) 299 ms (244-431)
WorkspaceSearchIndex.scanSymlinkEntries 551 ms (502-623) 207 ms (172-251)
WorkspaceSearchIndex.loadIgnoredPaths (git ls-files --ignored) 661 ms (630-761) 59 ms (52-87)
WorkspaceSearchIndex.scanDirectory spans in the first search 4,703 4,703

Later searches in the same project take 18-32 ms on Windows. The symlink scan is the largest single part of the first search on Linux and about 44% of it on Windows.

How to measure it again

  1. Check out the commit, then run vp i.
  2. Start a server on a fresh data dir from apps/server:
    T3CODE_DEV_SEED=0 node src/bin.ts serve --base-dir <fresh-dir> --port 39880 --host 127.0.0.1
    
  3. Save this as apps/server/zz-search.ts and run node zz-search.ts http://127.0.0.1:39880 <fresh-dir> <path-to-a-large-repo>:
import * as NodeServices from "@effect/platform-node/NodeServices";
import * as NodeWS from "@effect/platform-node/NodeWS";
import { ORCHESTRATION_PROTOCOL_VERSION, WsRpcGroup } from "@t3tools/contracts";
import * as Effect from "effect/Effect";
import * as Layer from "effect/Layer";
import { RpcClient, RpcSerialization } from "effect/rpc";
import { Socket } from "effect/socket";
import { execFileSync } from "node:child_process";

const [origin, home, cwd] = process.argv.slice(2);
const token = execFileSync(process.execPath, ["src/bin.ts", "auth", "session", "issue", "--token-only", "--base-dir", home], { encoding: "utf8" }).trim();
const { ticket } = await (await fetch(`${origin}/api/auth/websocket-ticket`, { method: "POST", headers: { authorization: `Bearer ${token}` } })).json();
await Effect.runPromise(Effect.gen(function* () {
  const protocol = yield* Layer.build(RpcClient.layerProtocolSocket().pipe(
    Layer.provide(Socket.layerWebSocket(`${origin.replace(/^http/, "ws")}/ws?orchestrationProtocol=${ORCHESTRATION_PROTOCOL_VERSION}&wsTicket=${encodeURIComponent(ticket)}`)),
    Layer.provide(RpcSerialization.layerJson), Layer.provide(NodeWS.layerWebSocketConstructor)));
  const client = yield* RpcClient.make(WsRpcGroup).pipe(Effect.provideContext(protocol));
  for (let i = 0; i < 3; i++) {
    const t = Date.now();
    yield* client["projects.searchEntries"]({ cwd, query: "manager", limit: 100 });
    console.log(`search_ms=${Date.now() - t}`);
  }
}).pipe(Effect.scoped, Effect.provide(NodeServices.layer)));
process.exit(0);
  1. Wait about 10 s so the trace flushes. Stop the server, then read the spans of the first search from <fresh-dir>/userdata/logs/server.trace.ndjson:
import json, sys
sp = [json.loads(l) for l in open(sys.argv[1]) if l.strip()]
root = min((s for s in sp if s["name"] == "ws.rpc.projects.searchEntries"), key=lambda s: int(s["startTimeUnixNano"]))
t0, te = int(root["startTimeUnixNano"]), int(root["endTimeUnixNano"])
kids = [s for s in sp if s["traceId"] == root["traceId"] and t0 <= int(s["startTimeUnixNano"]) <= te]
for s in kids:
    if s["name"] in ("ws.rpc.projects.searchEntries", "WorkspaceSearchIndex.loadIgnoredPaths", "WorkspaceSearchIndex.scanSymlinkEntries"):
        print(f'{(int(s["startTimeUnixNano"]) - t0) / 1e6:8.1f} {s["durationMs"]:8.1f} {s["name"]}')
print("scanDirectory:", sum(s["name"] == "WorkspaceSearchIndex.scanDirectory" for s in kids))

Repeat steps 2-4 with a new data dir each time, so every run starts cold.

Where the time goes

The code is in apps/server/src/workspace/WorkspaceSearchIndex.ts.

  • make waits for loadIgnoredPaths() and waitForIndexReady.
  • searchNames then calls listSymlinkEntries. On the first read this starts scanSymlinkEntries and waits for it in full.
  • scanSymlinkEntries asks fff for every indexed folder (listIndexedDirectories). It then runs scanDirectory on each one, which is NodeFSP.readdir(..., { withFileTypes: true }), 16 at a time. That is 4,703 folder reads to find 17 links.

Timeline of one cold Windows first search (ms from the request):

   1.4    8.1  WorkspaceSearchIndex.createFinder
   9.7  621.9  WorkspaceSearchIndex.loadIgnoredPaths   (git ls-files --others --ignored)
 631.7    1.1  WorkspaceSearchIndex.waitForIndexReady
 635.8  540.0  WorkspaceSearchIndex.scanSymlinkEntries  (4,703 scanDirectory spans)
1193.1         ws.rpc.projects.searchEntries ends

Why it is avoidable

The tree has just been walked by fff. The server then walks every folder again only to find the few symlinks fff skips, and the first search waits for that walk. The walk also starts only after the git ignored-paths listing, although it does not need that listing until it filters its result.


Found by the automated Windows performance sweep on f2d4df0367.

主要语言
TypeScript
星标
24.8k
派生
6.4k
平均合并
4 小时 29 分钟
30 天内合并 PR
488

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

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。

pingdotgg/t3code 的其他 Issue

查看 pingdotgg/t3code 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。