[GH-ISSUE #10] GCS mirror sync does not cover torch-index-stacks.json (or anything published outside the env build) #3

New issue

Open

opened 2026-10-26 22:22:45 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 22:22:45 +00:00

Owner

Originally created by @Kosinkadink on GitHub (Jul 25, 2026).
Original GitHub issue: https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/issues/10

The desktop app falls back to a public GCS mirror when R2 is unreachable (regions where the R2 edge is throttled, e.g. China):

Primary: https://desktop-assets.comfy.org/standalone-environments/
Mirror: https://storage.googleapis.com/comfy-desktop-public/standalone-environments/

The app's r2Mirror.ts assumes the mirror is a 1:1 copy of the R2 namespace, "kept in sync at each release". However, nothing in this repo performs that sync - both workflows (build-standalone-env.yml and the new publish-torch-index-stacks.yml from #9) only write to R2. The mirror is presumably synced by an outside process.

Problem

Once #9 lands, torch-index-stacks.json will be replaced in-place on R2 whenever the manifest changes - decoupled from env releases. If the mirror sync only runs at env-release time (or doesn't know about this object at all), mirror-dependent users get a stale PyTorch stack manifest indefinitely. That's worse than it sounds for two cases:

Withdrawals: pulling a known-bad stack via explicit removal would not reach mirror users.
New stacks: the whole point of the remote manifest is shipping stacks without an app release; mirror users would not receive them until the next unrelated env release (if ever).

Options

Add a GCS upload step to publish-torch-index-stacks.yml - upload the same file to the mirror bucket right after the R2 upload. Needs GCS credentials as repo secrets. Most direct; keeps the two objects atomically close in time.
Fold the manifest into the existing external mirror-sync process - whoever owns it adds this object (or switches to a full-prefix rsync of standalone-environments/, which also removes the per-object bookkeeping).
Scheduled sync workflow in this repo - an rsync-style job mirroring the whole prefix on a cron. Covers all current and future objects but adds propagation delay.

Option 1 (plus keeping the archive sync wherever it lives today) is probably the smallest correct fix; whoever owns the current mirror-sync process should weigh in.

Acceptance

A change to torch-index-stacks.json on main reaches both R2 and the GCS mirror without manual steps.
Documented (README or mirrors.md) who/what syncs the mirror, so the next published object doesn't repeat this gap.

Refs: #9, Comfy-Org/Comfy-Desktop#1248

Originally created by @Kosinkadink on GitHub (Jul 25, 2026). Original GitHub issue: https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/issues/10 The desktop app falls back to a public GCS mirror when R2 is unreachable (regions where the R2 edge is throttled, e.g. China): - Primary: `https://desktop-assets.comfy.org/standalone-environments/` - Mirror: `https://storage.googleapis.com/comfy-desktop-public/standalone-environments/` The app's `r2Mirror.ts` assumes the mirror is a 1:1 copy of the R2 namespace, "kept in sync at each release". However, nothing in this repo performs that sync - both workflows (`build-standalone-env.yml` and the new `publish-torch-index-stacks.yml` from #9) only write to R2. The mirror is presumably synced by an outside process. ## Problem Once #9 lands, `torch-index-stacks.json` will be replaced in-place on R2 whenever the manifest changes - decoupled from env releases. If the mirror sync only runs at env-release time (or doesn't know about this object at all), mirror-dependent users get a stale PyTorch stack manifest indefinitely. That's worse than it sounds for two cases: - **Withdrawals**: pulling a known-bad stack via explicit removal would not reach mirror users. - **New stacks**: the whole point of the remote manifest is shipping stacks without an app release; mirror users would not receive them until the next unrelated env release (if ever). ## Options 1. **Add a GCS upload step to `publish-torch-index-stacks.yml`** - upload the same file to the mirror bucket right after the R2 upload. Needs GCS credentials as repo secrets. Most direct; keeps the two objects atomically close in time. 2. **Fold the manifest into the existing external mirror-sync process** - whoever owns it adds this object (or switches to a full-prefix `rsync` of `standalone-environments/`, which also removes the per-object bookkeeping). 3. **Scheduled sync workflow in this repo** - an `rsync`-style job mirroring the whole prefix on a cron. Covers all current and future objects but adds propagation delay. Option 1 (plus keeping the archive sync wherever it lives today) is probably the smallest correct fix; whoever owns the current mirror-sync process should weigh in. ## Acceptance - A change to `torch-index-stacks.json` on `main` reaches both R2 and the GCS mirror without manual steps. - Documented (README or mirrors.md) who/what syncs the mirror, so the next published object doesn't repeat this gap. Refs: #9, Comfy-Org/Comfy-Desktop#1248

bsakaguchi commented

2026-10-26 22:22:46 +00:00

Author

Owner

@coderabbitai[bot] commented on GitHub (Jul 25, 2026):

🔗 Related PRs

Comfy-Org/ComfyUI-Standalone-Environments#9 - Add torch-index-stacks.json manifest with validated R2 publishing [open]

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

@coderabbitai[bot] commented on GitHub (Jul 25, 2026):  <details> <summary>🔗 Related PRs</summary> Comfy-Org/ComfyUI-Standalone-Environments#9 - Add torch-index-stacks.json manifest with validated R2 publishing [open] </details> --- <details> <summary> 🧪 Issue enrichment is currently in open beta.</summary> You can configure auto-planning by selecting labels in the issue_enrichment configuration. To disable automatic issue enrichment, add the following to your `.coderabbit.yaml`: ```yaml issue_enrichment: auto_enrich: enabled: false ``` </details> 💬 Have feedback or questions? Drop into our [discord](https://discord.gg/coderabbit)!

bsakaguchi referenced this issue

2026-10-26 22:22:52 +00:00

[PR #3] [MERGED] fix: pin pytorch package versions in all requirements files #6