Files
setup-buildx/README.md
T
Anders OlssonandClaude Opus 5 17d77ace6a fix: default to the buildkit in the caller's own namespace
The default was `tcp://buildkit.gitea.svc.cluster.local:1234`, which
names aceofbase's namespace explicitly. That was correct while aceofbase
was the only cluster running CI, and resolves to nothing from
planet-express — so every workflow moved to the new dink-backed runner
had to override it, and forgetting produced a DNS error from inside
buildx that named neither this action nor the cluster it pointed at.

A BARE Service name resolves in the job container's own namespace, which
is where its runner's buildkit lives on both clusters: `gitea` on
aceofbase, `gitea-runner` on planet-express. One default, correct on
both, and correct for a cluster nobody has built yet.

Verified before changing rather than assumed: on aceofbase the buildkit
Service is in `gitea`, and gitea-act-runner's Role is namespace-scoped
to `gitea`, so its job pods land in the same namespace as the Service.
Existing consumers therefore see no change.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01C1tzpqzv37DDbtpPTxHsyA
2026-08-18 06:25:35 +02:00

70 lines
2.8 KiB
Markdown

# setup-buildx
Drop-in replacement for `docker/setup-buildx-action@v3` that defaults to the
`buildkit` service in the calling job's own namespace.
Use this whenever a Gitea Actions workflow needs to build images. It removes
the boilerplate of declaring a remote driver every time, and ensures every
build reuses the cluster's BuildKit instance (persistent 100 GiB Longhorn
PVC + tuned GC) for layer caching across runs.
## Usage
```yaml
jobs:
build:
runs-on: aceofba-cluster
container:
image: ghcr.io/catthehacker/ubuntu:act-22.04
steps:
- uses: actions/checkout@v4
- uses: https://git.aceofba.se/infra/setup-buildx@v1
- uses: docker/build-push-action@v6
with:
context: .
file: Dockerfile
push: true
tags: git.aceofba.se/${{ github.repository }}:latest
cache-from: type=gha
cache-to: type=gha,mode=max
```
## Inputs
| Name | Default | Description |
|------|---------|-------------|
| `endpoint` | `tcp://buildkit:1234` | BuildKit TCP endpoint. A bare Service name, so it resolves in the job's own namespace. Override only if you run a private buildkit somewhere else. |
| `version` | `latest` | buildx version, passed through to the upstream action. |
## Why
`docker/setup-buildx-action@v3` defaults to `driver: docker-container`, which
requires a working Docker daemon inside the job container. The act-runner
uses Kubernetes pod-per-job hooks — there is no dockerd in those pods. The
cluster does run a standalone BuildKit Service (`buildkit-service` Helm
chart, port 1234) which buildx can talk to via its `remote` driver. This
action wires that in by default.
`cache-from`/`cache-to: type=gha` also works out of the box: the act-runner
image (christopherhx/gitea-actions-runner) ships its own embedded Actions
Cache / Results API and injects `ACTIONS_CACHE_URL` / `ACTIONS_RUNTIME_TOKEN`
into the job container. That cache is persisted on the runner's `/data` PVC,
so it survives pod restarts. There is no separate cache-server deployment to
configure.
## Why the endpoint is a bare Service name
A job container resolves a bare Service name in its own namespace, which is
where its runner's buildkit lives on every cluster here — `gitea` on
aceofbase, `gitea-runner` on planet-express. One default is therefore correct
on both, and stays correct for a cluster nobody has built yet.
Until 2026-08-18 it defaulted to `tcp://buildkit.gitea.svc.cluster.local:1234`,
naming aceofbase's namespace explicitly. That was fine while aceofbase was the
only cluster running CI. It resolved to nothing from planet-express, so every
workflow moved there had to pass `endpoint:` by hand, and forgetting produced a
DNS error from inside buildx that named neither this action nor the cluster it
was pointing at.