Build and push container images with Buildah on the quadronix ARC runners (arc-azrtydxb-amd64).
A drop-in for docker/setup-buildx-action + docker/build-push-action, without BuildKit, without
a registry layer cache to configure, and with multi-arch builds that never use emulation.
- uses: actions/checkout@v7
- uses: docker/login-action@v4 # or azrtydxb/ant-setup-action with docker-login: true
with: { registry: ghcr.io, username: ${{ github.actor }}, password: ${{ secrets.GITHUB_TOKEN }} }
- uses: azrtydxb/image-build-action@v1
id: img
with:
context: .
file: build/Dockerfile
platforms: linux/amd64,linux/arm64
push: true
tags: |
ghcr.io/azrtydxb/app:${{ github.sha }}
ghcr.io/azrtydxb/app:mainEach runner pod of arc-azrtydxb-amd64 has a rootless podman system service sidecar
(quay.io/podman/stable, BUILDAH_LAYERS=true, native overlay, BUILDAH_ISOLATION=chroot) and the
env CONTAINER_HOST=unix:///run/podman/podman.sock. The action downloads a pinned, checksum-verified
static podman-remote client per job (no custom runner image), then:
podman-remote build --layers --platform <list>(with--manifest <unique name>for several platforms,-t <unique name> --iidfilefor one);- pushes to every tag with
podman-remote manifest push --all(several platforms) orpodman-remote push(one), using the job's credentials; - untags its local names again.
The image store (/data/buildah/storage) and the RUN --mount=type=cache directory
(/data/buildah/cache) are shared by all runners on the host, so a layer built by any earlier job
is a cache hit for the next job on any pod (warm builds: seconds instead of minutes). Base images are
pulled through the home Ant (ant.kw.watteel.lab) via the sidecar's registries.conf.
A nightly CronJob (quadronix/arc/buildah-gc.yaml in azrty-infra) prunes the store when it grows past
a limit, and never while a build runs (builds hold a shared flock).
| Input | Default | Meaning |
|---|---|---|
context |
. |
Build context directory |
file |
<context>/Dockerfile |
Dockerfile path |
platforms |
linux/amd64 |
Comma-separated; more than one builds a manifest list |
tags |
One per line or comma-separated; required with push |
|
push |
false |
Push every tag using the job's registry credentials |
load |
false |
Single platform: also docker load the image into the runner's dind |
build-args |
KEY=VALUE per line |
|
labels |
KEY=VALUE per line |
|
target |
Target stage | |
secrets |
id=VALUE per line, for RUN --mount=type=secret,id=... |
|
secret-files |
id=PATH per line |
|
no-cache |
false |
--no-cache |
pull |
false |
--pull=always |
auth-file |
$DOCKER_CONFIG/config.json or ~/.docker/config.json |
Credentials; where docker/login-action and ant-setup-action write theirs |
sbom |
false |
With push: SPDX SBOM (syft 1.54.1) attached as a keyless cosign attestation to the pushed image, or to each platform manifest of a list |
provenance |
false |
With push: SLSA v1 provenance attached as a keyless cosign attestation to the pushed digest (the index for several platforms) |
insecure-registries |
Plain-HTTP or untrusted-TLS registries (localhost:5001): pushes to those use --tls-verify=false, cosign/syft/crane get their insecure flags |
|
podman-version |
6.1.3 |
Only versions in the checksum table of build.sh |
Outputs: platform-digests (one os/arch digest line per platform manifest of a pushed list), digest (of the pushed image or manifest list; empty without push), imageid, metadata
(JSON with containerimage.digest, containerimage.config.digest, image.name).
Remove these, they have no use with this action:
docker/setup-buildx-actionandazrtydxb/.github/actions/setup-builder(no builder to create);cache-from/cache-to(type=registry fromant-setup-action'sbuild-cacheoutputs, or type=gha): the shared store is the cache. Also dropbuild-cache: "true"fromant-setup-actionif the job builds nothing else with buildx;builder,github-token,no-cache-filters,outputs(not supported; BuildKit-only).sbom: trueandprovenance: mode=maxkeep working: they map to the inputs below (any value butfalseturns them on).
Keep docker/login-action (or ant-setup-action with docker-login: true): the action reads
~/.docker/config.json. Everything else maps by name.
Before:
- uses: docker/setup-buildx-action@v4
- id: ant
uses: azrtydxb/ant-setup-action@v1
with: { docker-login: true, build-cache: true }
- uses: docker/build-push-action@v7
with:
context: .
file: build/docker/api.Dockerfile
platforms: linux/amd64,linux/arm64
push: true
provenance: false
tags: ghcr.io/azrtydxb/api:${{ github.sha }}
cache-from: ${{ steps.ant.outputs.cache-from }}
cache-to: ${{ steps.ant.outputs.cache-to }}After:
- uses: azrtydxb/ant-setup-action@v1
with: { docker-login: true }
- uses: azrtydxb/image-build-action@v1
with:
context: .
file: build/docker/api.Dockerfile
platforms: linux/amd64,linux/arm64
push: true
tags: ghcr.io/azrtydxb/api:${{ github.sha }}steps.<id>.outputs.digest is the same as docker/build-push-action's.
sbom: true and provenance: true replace BuildKit's attestations. After the push the action installs
pinned, checksum-verified syft and cosign and attaches, keyless (GitHub OIDC identity, Sigstore
Fulcio/Rekor), in the registry next to each repository of tags:
- an SPDX JSON SBOM (
--type spdxjson) on the image, or on each platform manifest of a manifest list; - a SLSA v1 provenance (
--type slsaprovenance1: workflow, ref, commit, run) on the pushed digest, i.e. the index for several platforms.
The job needs permissions: id-token: write (the action fails with a clear message without it); pushing
to ghcr.io also needs packages: write. Verify with:
cosign verify-attestation --type slsaprovenance1 \
--certificate-identity-regexp 'https://github.com/azrtydxb/.*' \
--certificate-oidc-issuer https://token.actions.githubusercontent.com ghcr.io/azrtydxb/app@sha256:...Attestations are stored as registry artifacts (cosign's sha256-<digest>.att tags), which any registry
that stores tags and manifests accepts (ghcr.io, OCIR and Ant's oci-hosted feed need no referrers API).
GitHub's own attestation store (actions/attest-build-provenance) is not used: it is not available for
private repositories on the org's plan. They need no attestations: write permission.
attest-only: true skips the build and attaches the attestations to digest in each repository of
tags, for example the index a later job created with docker buildx imagetools create:
- uses: azrtydxb/image-build-action@v1
with:
attest-only: true
tags: ghcr.io/azrtydxb/app:${{ github.sha }}
digest: ${{ steps.index.outputs.digest }}
provenance: true- Runner:
runs-on: arc-azrtydxb-amd64(or its labels), a step directly in the runner container. Acontainer:job runs in another pod that cannot reach the socket; the action then fails with a clear message. Build in a separate job withoutcontainer:. GitHub-hosted runners and the arm64 OKE runners have no sidecar either. - Cross-arch without emulation: QEMU is not allowed. For
linux/arm64on this amd64 host the Dockerfile must cross-compile: build stagesFROM --platform=$BUILDPLATFORM, use$TARGETOS/$TARGETARCH(GOOS/GOARCH,cargo --target, ...), and the final stage must have noRUN(useCOPY --from). ARUNin a foreign-arch stage fails withexec format error. - Dockerfile syntax: Buildah is not BuildKit.
RUN --mount=type=cache|secret|bind|tmpfswork;RUN --network,ADD --checksum/--link/--keep-git-dir,COPY --exclude, heredocs inRUN(supported) are the ones to check;# syntax=docker/dockerfile:1is ignored. - Image names are unique per job (the store is shared); do not rely on a local name between steps.
Use
load: true(single platform) to get the image into dind, or pull what you pushed. - Secrets:
RUN --mount=type=secret,id=xreads/run/secrets/x. Never pass secrets as build-args. - The build runs under a shared lock so the store GC cannot prune during it; an unusually long build only delays the GC.
HOME-based auth:docker loginmust have run in an earlier step of the same job.
.github/workflows/test.yml builds a tiny multi-arch Dockerfile on arc-azrtydxb-amd64, pushes it
to ghcr.io/azrtydxb/image-build-action-test, checks both architectures in the manifest list,
a second run for cache hits, and a single-platform load. Consumers pin @v1; the v1 tag moves
with compatible changes, v1.x.y tags are immutable.