8. Porting to GitHub Actions¶
Why this matters¶
Pipelines outlive platforms. Knowing the concepts behind this repo's
GitLab pipeline — ordering, change detection, manual gates, OIDC signing —
means you can rebuild it on any CI system. .github/workflows/ contains a
working GitHub Actions port of the entire pipeline; this lesson is the
mapping table that explains it.
The concept map¶
| GitLab CI | GitHub Actions |
|---|---|
.gitlab-ci.yml + include: local: |
.github/workflows/*.yml — workflow_call reusable workflows |
stages: |
Nothing — ordering comes entirely from needs: (jobs parallelize by default) |
workflow.rules (MR vs branch vs tag) |
on: pull_request / on: push triggers |
rules: changes: |
paths: on the trigger, or dorny/paths-filter for per-job change detection |
image: per job |
container: per job, or tool-setup actions (setup-java, setup-helm) |
CI_REGISTRY_* predefined vars |
ghcr.io + GITHUB_TOKEN — nothing to configure |
CI_COMMIT_SHORT_SHA |
${{ github.sha }} (full SHA; ${GITHUB_SHA::7} in shell for short) |
cache: |
actions/cache, or cache: maven inside setup-java |
artifacts: / reports:junit |
actions/upload-artifact + dorny/test-reporter |
when: manual |
environment: with required reviewers (Settings → Environments) |
resource_group: prod |
concurrency: prod |
id_tokens: (OIDC) |
permissions: id-token: write |
| Secret-Detection template | gitleaks/gitleaks-action |
allow_failure: true |
continue-on-error: true |
| kaniko executor image | docker/build-push-action — runners have a daemon, no kaniko needed |
| protected CI variable | environment-scoped secret |
pages: job (GitLab Pages) |
actions/upload-pages-artifact + actions/deploy-pages |
File-for-file¶
.gitlab-ci.yml → .github/workflows/ci.yml
.gitlab/ci/test.yml → .github/workflows/reusable-test.yml
.gitlab/ci/build-kaniko.yml → .github/workflows/reusable-build.yml
.gitlab/ci/security.yml → .github/workflows/reusable-security.yml
.gitlab/ci/package-chart.yml → .github/workflows/reusable-package.yml
.gitlab/ci/deploy.yml → .github/workflows/reusable-deploy.yml
.github/workflows/rollback.yml
(a pages: job, if one existed) → .github/workflows/docs.yml
docs.yml is a separate workflow, not a caller in ci.yml — on purpose.
The docs site has its own change filter (docs/**), its own concurrency
group (pages allows one deployment at a time), and its own permissions
(pages: write). Not everything belongs in one pipeline; trigger-scoped
paths: is the native way to say "only run when these files change."
The .deploy-base anchor becomes a single parameterized reusable
workflow: deploy-dev and deploy-prod are two call sites passing
different typed inputs: — the same pattern as GitLab CI/CD components.
The three genuinely different things¶
1. No stages — needs is the only ordering¶
GitLab runs stage-by-stage; a job with no needs waits for the whole
previous stage. GitHub runs everything in parallel unless needs says
otherwise. The port expresses the whole DAG explicitly — including the
GitLab optional: true semantics, which become an always() + result
check:
if: >-
always() &&
needs.changes.outputs.build == 'true' &&
(needs.test.result == 'success' || needs.test.result == 'skipped')
Without always(), a job whose need was skipped is itself skipped —
that's the gotcha behind every "optional needs" translation.
2. rules: changes: is per-job; paths: is per-workflow¶
GitHub's native paths: filter decides whether the whole workflow runs.
Per-job granularity needs a fan-out job (dorny/paths-filter) that
exports flags the downstream jobs read via needs.changes.outputs.*.
3. Manual gates live on the environment, not the job¶
GitLab's when: manual draws a button inside the pipeline. GitHub instead
pauses a job on an environment: that has required reviewers
configured — set them under Settings → Environments → prod. Same
outcome (pipeline waits for a human), different plumbing.
The rollback:prod job has no in-pipeline equivalent, so it becomes a
standalone workflow_dispatch workflow — run it on demand from the
Actions tab. It keeps environment: prod, so the same approvals and
secrets apply.
What you must configure once¶
- Secrets: add
KUBECONFIG_DATA(base64 kubeconfig). Scope it to thedev/prodenvironments to reproduce protected-variable semantics. - Environments: create
devandprod; add required reviewers toprodto get thewhen: manualgate. - Nothing for the registry:
GITHUB_TOKEN+ghcr.ioworks out of the box. On PRs from forks the token is read-only — pushes to GHCR will fail there, same as a fork MR failing to push to a private GitLab registry.
What doesn't port¶
- The dedup rule. GitLab's "skip the branch pipeline when an MR covers
the commit" has no native twin — the port sidesteps it by triggering
pushonmainonly (PR branches are covered bypull_request). - Merged-results vs merge commit: subtle, but both platforms test the post-merge result by default on MR/PR events, so behavior matches.