voidzero-dev / voidzero-dev/vite-plus

vp migrate hangs indefinitely (0% CPU) when the project's pinned pnpm deadlocks during install

Abierto
#2,020 1 comentario 0 reacciones 1 asignado Reclamado por @fengmk2 Ver en GitHub
Lenguaje dominante
Rust
Estrellas
5.8k
Forks
262
Merge medio
1 d 34 min
PR fusionados (30 d)
135

Descripción

## Summary

`vp migrate` can hang **indefinitely at 0% CPU** (no output, no progress, no timeout) on some projects. The hang is in the dependency install that migrate spawns: the project's **pinned pnpm 9.x** deadlocks on an internal worker pipe. Because `vp migrate` has no timeout or hang detection, the command never returns and must be killed manually.

Discovered while smoke-testing the v0.2.2 prerelease (PR #2016, registry-bridge commit `1d7ba811f0e456640b0e761e8e8acf77d1ad3251`) across ~64 real-world projects: 2 of them hang every time, the other ~62 migrate fine.

## Affected projects (100% reproducible)

| Project | Pinned pnpm | Node |
| --- | --- | --- |
| `anywidget` | `pnpm@9.10.0` | 24.18.0 |
| `vitepress-carbon` | `pnpm@9.9.0` | 22.18.0 (`.node-version`) |

Both are pnpm monorepos. ~62 other projects in the same run (mostly newer pnpm) migrated successfully through the same registry bridge, so this is specific to these projects' resolution + their old pinned pnpm.

## Symptom

`vp migrate` prints nothing after startup and never returns:

```
Running vp migrate in /Users/.../anywidget
<-- hangs here forever; killed after 20 min every time -->
```

- The `vp` process and its `node`/`pnpm` children all sit at **0.0% CPU** (idle deadlock, not a busy loop).
- Not a stdin prompt: reproduces identically with `stdin` redirected from `/dev/null`.
- Not concurrency: reproduces when run solo with nothing else installing.

## Root-cause evidence

The hung process is the install that migrate spawns:

```
node .../pkg-pr-new//package_manager/pnpm/9.10.0/pnpm/bin/pnpm.cjs install --no-frozen-lockfile
```

**Stack sample of the hung node/pnpm main thread** (`sample `):

```
node::SpinEventLoopInternal
uv_run
uv__io_poll
kevent <-- main thread parked in the event loop waiting on an fd
```

**`lsof` of the hung process** shows **no TCP/network connections — only internal PIPE fds**:

```
node ... 4 PIPE 0x...->0x...
node ... 5 PIPE 0x...->0x...
node ... 12 PIPE 0x...->0x...
node ... 13 PIPE 0x...->0x...
```

So it is **not** stalled on the registry/network. pnpm's main process is blocked in the event loop waiting on an internal worker pipe that never signals, i.e. an internal pnpm worker-IPC deadlock (same class as the pnpm 11.0.6 "hangs after Done" bug that was fixed in 11.5.1, but here on pnpm 9.x).

## Reproduction

```bash
# via the pkg.pr.new bridge migrate helper
ALLOW_DIRTY=1 .github/scripts/test-pkg-pr-new-migrate.sh \
1d7ba811f0e456640b0e761e8e8acf77d1ad3251 /path/to/anywidget --no-interactive
# ...or plainly: cd anywidget && vp migrate
# hangs at "Running vp migrate ..." with 0% CPU; sample/lsof the pnpm child as above.
```

## Impact

- `vp migrate` never completing is a poor failure mode: no error, no timeout, silent 0%-CPU hang; a CI or a user waits forever.
- In an automated fleet run it also leaves **orphaned hung pnpm processes** behind (observed several, up to 1h+ old) that must be reaped manually.

## Suggested directions

1. **Add a timeout / hang detection to the install step of `vp migrate`** so it fails loudly (with the pnpm command + partial output) instead of hanging silently forever. This alone converts the worst part (silent infinite hang) into an actionable error.
2. Consider running the migrate install with **vp's managed/bundled pnpm** (newer) rather than the project's old pinned `pnpm@9.x` that deadlocks, or expose a flag to opt into that.
3. Investigate the underlying pnpm 9.x worker-pipe deadlock on these two projects' graphs (likely reproducible with a plain `pnpm@9.10.0 install` on `anywidget`).

_Found during the v0.2.2 real-world test; see https://github.com/voidzero-dev/vite-plus/pull/2016 for the prerelease build under test._

Guía de contribución

Abrir la guía de contribución

Evaluación

Este issue todavía no se ha evaluado.

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.