[Coly010]

Summary

In stack mode, Edge Functions that boot at the same time can fail with 503 BOOT_ERROR. The logs show Failed caching npm package … File exists (os error 17), or No such file or directory when reading a cached package file. This sets noModuleCache: false when the stack bootstrap creates user workers, which is what classic functions serve already does.

What's going on

noModuleCache: true puts the worker's Deno cache into reload mode. Each worker builds its own npm cache, so nothing is shared between workers, and each one downloads every npm package it imports again. When that package is already in the shared DENO_DIR, Deno extracts it again in place, over the live copy, deleting and recreating each file. Two workers doing that at the same time collide, and a worker that reads a file in the middle of it finds it missing.

flowchart TD
  boot["Worker boots and imports npm:pkg"] --> flag{"noModuleCache"}
  flag -- "false" --> cached{"pkg in the cache?"}
  cached -- "yes" --> reuse["Reuse it"]
  cached -- "no" --> temp["Extract to a temp dir,<br/>then rename into place"]
  flag -- "true" --> cached2{"pkg in the cache?"}
  cached2 -- "no" --> temp
  cached2 -- "yes" --> inplace["Extract again in place,<br/>over the live copy"]
  inplace --> fail["Concurrent boots collide<br/>503 BOOT_ERROR"]

Loading

sequenceDiagram
  participant A as Worker A
  participant C as Shared Deno cache
  participant B as Worker B
  A->>C: create pkg/README.md (exists, so delete it)
  B->>C: create pkg/README.md (succeeds)
  A->>C: create pkg/README.md
  C-->>A: File exists (os error 17)
  Note over A: 503 BOOT_ERROR

Loading

This explains the odd parts of the report:

• A package that isn't cached yet goes through the temp dir and rename, which is safe. That's why a cold first round often passes, and why a warm-up request doesn't help.
• Native keeps its cache across stop and start, so it fails much more often than Docker, whose new container starts with an empty cache.
• forceCreate: true boots a new worker for every request, so every request goes down this path. I've left it as it is, because it's what makes edits show up without a restart.

Against Edge Runtime directly, with only this flag changed, 40 of 64 concurrent boots failed with it on and none of 112 failed with it off.

What changes

• npm packages and remote modules are reused from the cache instead of being downloaded and extracted on every boot, so boots are quicker too.
• An exact version that isn't cached yet still resolves, because Deno fetches package info again when resolution fails. Floating versions like npm:x@latest stay on the cached version until the cache resets (a new container on Docker, stack destroy on native). That matches classic functions serve and deno run.
• Local function code is still read from disk on every boot, so edits show up straight away.
• A new integration test serves a package from a local npm registry, warms the cache, then boots several workers at once. It checks that they all succeed and that the package is downloaded only once.

Earlier bootstraps and classic functions serve use false. If true was there for something I've missed, I'm happy to talk it through.

Linked issue

Closes #7321 (CLI-2850)

[github-actions]

🤖 AI Review

Both independent reviews reported zero findings. Inspection of the PR code found no additional issues: the cache setting, generated bundle, and regression tests are consistent. No fundamental conflict with the pending next-branch work was identified. Tests were not run during adjudication.

Findings

No issues found.

Stats

Claude findings: 0 · Codex findings: 0 · Confirmed: 0 · Refuted: 0 · Uncertain: 0

---

Models: claude-opus-5-5 + gpt-6.1-sol · Trigger: auto · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.