CacheFlow vs the GitHub Actions cache

As of

GitHub’s Actions cache is built into every repository and needs no setup on GitHub-hosted runners. CacheFlow serves the same Actions cache protocols (v2 twirp and v1) and the Actions artifact service from your own Cloudflare account, so stock actions/cache, setup-node, Swatinem/rust-cache@v2 and actions/upload-artifact use it with no workflow change, on runners whose environment points ACTIONS_RESULTS_URL at CacheFlow.

What CacheFlow adds is control: where the bytes live, a provenance record per entry, a policy that makes chosen jobs restore only what a trusted build saved, and a record of every restore, miss and save. What it does not add is speed. Bytes come from Cloudflare R2 over the internet, and we have not measured a restore-speed advantage over GitHub’s cache.

At a glance

Aspect GitHub Actions cacheCacheFlow
Setup [1] Built in: actions/cache uses it with no configuration. Set ACTIONS_RESULTS_URL once in the runner environment; workflows stay unchanged. Needs runners whose environment you control. On GitHub-hosted runners the same workflow uses GitHub’s cache.
Who operates it GitHub. You. It runs on your Cloudflare account: bytes in R2, ownership records in D1, a read path in KV, sessions in Durable Objects.
Storage limit [1] 10 GB per repository by default. Enterprise, organization or repository admins can raise it, and storage beyond 10 GB incurs additional cost. User-owned repositories can configure up to 10 TB. Per-namespace size limits that you set. Storage is R2 in your own account.
Eviction [1] Entries not accessed in over 7 days are removed. Over the limit, the least recently accessed entries go first. Namespace size limits only. No time-based eviction.
Branch scoping [1] A run restores caches from its own branch or the default branch; a pull request run can also restore from its base branch, including base branches of forked repositories. Caches saved by a pull request are scoped to its merge ref. Drop-in jobs follow GitHub’s own cache grant from the runtime token, so they read and write exactly the scopes they would on GitHub, including the repository’s real default branch. OIDC (opt-in) jobs write to their own ref scope and read their own ref, then the pull request base, then a configured list of fallback branches (main and master unless an operator adds others, such as trunk).
Low-trust events [1] Runs triggered by other events that resolve to the default branch get read-only access to the default branch’s caches. Drop-in jobs follow GitHub’s grant, as on GitHub. On OIDC (opt-in) jobs, saves from events outsiders can trigger on a protected branch, such as pull_request_target or issue_comment, go to a quarantined scope; pull_request events stay isolated to refs/pull/N/merge.
Entry provenance Not covered in the cited reference. Each entry records event, ref, commit and workflow (full on the OIDC path; drop-in records the event as unknown). A per-tenant trusted-consumer policy makes the jobs it covers restore only entries from trusted producers: push, schedule and manual runs on the protected branch. Without a policy, restores are ordinary branch-scoped lookups.
Rate limits [1] Up to 200 uploads and 1,500 downloads per minute per repository. Not published.
Telemetry Not covered in the cited reference. Restores (exact and prefix), misses and saves are recorded with bytes for the Actions cache and artifacts. The telemetry records Actions cache keys and artifact names only as a keyed per-tenant tag, never as written (the cache itself stores keys as usual). A per-repository summary is operator-only; there is no tenant view yet.
Artifacts A separate GitHub feature, not compared here. Serves the Actions artifact service too, so upload-artifact and download-artifact use it. Per-tenant retention with a 30-day ceiling.
Restore speed Not compared. No measured advantage. Bytes come from R2 over the internet.

Where CacheFlow is the better fit

  • You already run self-hosted runners and can set one environment variable on them.
  • You want cache and artifact bytes in a Cloudflare account you own, with size limits you set.
  • Release or deploy jobs should restore only what a trusted build saved: a trusted-consumer policy limits them to entries from push, schedule and manual runs on the protected branch, using the provenance recorded on OIDC saves.
  • You want every restore, miss and save, with its size, recorded for the Actions cache and artifacts, alongside Turborepo uploads and downloads, in one place.

Where GitHub’s Actions cache is the better fit

  • Your jobs run on GitHub-hosted runners. CacheFlow’s drop-in needs a runner environment you control; on hosted runners the workflow keeps using GitHub’s cache anyway.
  • You do not want to operate anything. GitHub runs its cache for you; CacheFlow runs on your Cloudflare account and is yours to operate.
  • Your caches fit in the default 10 GB, or paying GitHub for a larger limit is simpler than running a service.
  • Restore speed is the goal. CacheFlow has no measured speed advantage over GitHub’s cache.
  • Open-source projects with many fork pull requests, where GitHub’s base-branch restore rules for forks already do what you need.

How to switch

Point the runner environment at CacheFlow once, per runner pool. Workflows do not change, and the same workflow still uses GitHub’s cache on GitHub-hosted runners.

# runner environment: set once, per runner pool
ACTIONS_RESULTS_URL=https://cache.cacheflow.sh/

# your workflows: unchanged
- uses: actions/cache@v4
- uses: Swatinem/rust-cache@v2
- uses: actions/upload-artifact@v4

Sources

Claims about other products come from their public documentation, retrieved on . Claims about CacheFlow describe what ships today; planned work is marked as planned.

  1. GitHub Docs: Dependency caching reference: https://docs.github.com/en/actions/reference/workflows-and-actions/dependency-caching (retrieved 2026-09-28)