Files
devops-lib-gcp/README.md
T
Mukul SharmaandClaude Opus 5 6d743cbe48 Make the registry hostname configurable and drop the CA mount
buildDocker.groovy hardcoded harbor.35.238.248.203.nip.io as the push
target, so the registry could not move without editing this shared library
and every consumer moving in the same commit. It now reads
config.harbor_registry, whose default lives in homelabPipeline.groovy
beside harbor_project and every other key.

The stage errors rather than defaulting when the value is missing. Carrying
a second copy of the literal would leave two defaults free to disagree, and
an unset value would otherwise build an image named "null/<project>/<repo>"
— which docker accepts as a hostname and then fails to resolve, pointing
nowhere near the cause.

The dind pod no longer mounts the registry CA. That mount existed because
the registry was a nip.io name, which no public CA will issue for, so
cert-manager signed Harbor from a private CA; the node pool was told to
trust it for pulls, but a push comes from dockerd inside the build pod,
which has its own trust store. harbor.infra.deployshed.com carries a Let's
Encrypt certificate that both already trust, so the mount, its volume and
the whole arrangement go away rather than being repointed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LEsTefWWifp4ikvhHF5s6N
2026-09-17 09:31:32 +05:30

124 lines
6.8 KiB
Markdown

# devops-lib (GKE)
Jenkins Shared Library for the CI/CD pipeline on the GKE cluster. The GCP
counterpart of the homelab repo of the same name, and a copy rather than a
shared repo because several values here are cluster-specific in ways that
would break the other cluster if crossed over.
Register it in Jenkins under the name `devops-lib`, exactly as on the
homelab. Consuming repos then need no change at all — the same two-line
Jenkinsfile works on either cluster, and which library it resolves to is a
property of the Jenkins it runs on.
**What differs from the homelab copy**, all of it a consequence of GKE
being a real cloud rather than one VM:
- **The registry hostname** is `harbor.infra.deployshed.com`, in the push
target (`harbor_registry`, see the table below) and in all five fallback
Dockerfiles.
- **Harbor speaks TLS, with a public certificate.** The homelab's dind
passes `--insecure-registry`; here nothing is needed at all, because
Let's Encrypt issues for the real domain and dockerd already trusts it.
- **`helm_repo_url` uses cluster DNS**, since the clone happens inside a
build pod. The homelab points it at an ingress hostname.
- **`build-tools` is not in this repo.** It lives in
`devops-base-images-gcp` alongside the mirrored base images, and is
referenced here only by tag.
Everything else — the stage flow, the fallback templates, the hooks
contract — is unchanged. The library was already adapted from a much
larger, company-wide one; see git history for what was removed.
## Using it in a service repo
The entire Jenkinsfile is 2 lines:
```groovy
@Library('devops-lib') _
homelabPipeline(repo_name: 'my-service')
```
`repo_name` is the only required key. Everything else has a sensible
default — override any of them by passing extra keys to `homelabPipeline`,
or by committing a `config.yaml` to the service repo's own root (merged
in after checkout; repo-committed values win over the Jenkinsfile call).
| Key | Default | Notes |
|---|---|---|
| `service_name` | `repo_name` | Second path segment under `devops-helm-charts/values/` |
| `argo_app_name` | `repo_name` | Must match the ArgoCD Application's `metadata.name` |
| `harbor_registry` | `harbor.infra.deployshed.com` | The registry hostname images are pushed to and pulled from. Must be spelled identically here, in the `dockerconfigjson` `auths` key, and in Harbor's `externalURL` — docker matches both stored credentials and TLS trust by exact hostname, so a mismatch fails as `unauthorized` rather than as a name problem |
| `harbor_project` | `apps-registry` | Must already exist in Harbor, and be public unless you also wire an `imagePullSecret` — the app values assume anonymous pull. A push to a missing project fails as `unauthorized: project <name> not found` |
| `helm_repo_url` | `devops-helm-charts-gcp`, over cluster DNS | `http://gitea-http.gitea.svc.cluster.local:3000/gitadmin/…` — pod-to-pod, so it never leaves the cluster and comes back through the ingress |
| `image_tag_yq_path` | `.deployment.image.tag` | **Override this if the app's chart isn't `1.0.0`** — e.g. `sts-2.0.0` uses `.podtemplate.image.tag` instead. Getting this wrong doesn't fail loudly: `yq -i` creates the path if missing rather than erroring, silently leaving the real field un-bumped. |
| `dockerBuildVersion` | none | Only read when the repo has **no Dockerfile of its own** — picks a fallback template (see below). No default; either ship a Dockerfile or set this. |
## Pipeline stages
`checkOut → loadConfig → runHooks(pre_build) → buildDocker →
runHooks(post_build) → updateHelmTag → syncArgoApp → notify`, all inside
a `podTemplate` (`resources/org/homelab/dind-pod.yaml`) via
`node(POD_LABEL) { ... }`.
- **`loadConfig`** — if the repo has a `config.yaml` at its root, its
keys are merged into the pipeline config (repo values win).
- **`runHooks`** — reads `config.yaml`'s `hooks.pre_build`/`hooks.post_build`
lists, each `{name, script, interpreter, requirements, blocking,
timeout_seconds}`. Blocking by default; `blocking: false` demotes a
failure to advisory (log + continue). Script paths must be
repo-relative (no `..`, no absolute paths).
- **`buildDocker`** — uses the repo's own `Dockerfile` if present;
otherwise renders one from `resources/com/homelab/<lang>-Dockerfile`
based on `dockerBuildVersion` (e.g. `go-1.22`, `node-20`,
`python-3.12`, `java-21`, `php-8.3`). All fallback templates pull base
images from Harbor's `base-images` project (mirrored via the separate
`devops-base-images-gcp` repo), not Docker Hub directly. Only versions
actually mirrored there resolve — an unmirrored tag fails the build
rather than silently falling back to Docker Hub.
- **`updateHelmTag`** — clones `devops-helm-charts`, bumps the image tag
via `yq` at `image_tag_yq_path`, commits, pushes to `main`.
- **`syncArgoApp`** — calls the ArgoCD REST API to sync `argo_app_name`.
## Adding a new language's fallback template
1. Add the base image to `devops-base-images/images.txt`, re-mirror it
into Harbor.
2. Add `resources/com/homelab/<lang>-Dockerfile`, parametrized by
`${version}` (rendered via `constructTemplate.groovy`'s
`SimpleTemplateEngine` wrapper).
3. Add a case for it in `buildDocker.groovy`'s `templates` map.
## Build-tools image
The `docker-cli` container runs
`harbor.infra.deployshed.com/base-images/build-tools:1`, which bakes in
git, yq, bash, python3 with pip and venv, and curl, so nothing is installed
on demand on every build.
**It is not built here.** The Dockerfile lives in `devops-base-images-gcp`,
next to the mirrored base images, because it is the same kind of artefact:
built by hand, occasionally, and pushed to Harbor. A Jenkins job could not
build it anyway — it is the image Jenkins builds *in*.
`dind-pod.yaml` pins the tag, so rebuilding the image rolls nothing out
until that pin is bumped. Bump the tag rather than overwriting one.
## Registry trust
Nothing to configure. Harbor is reached at `harbor.infra.deployshed.com`,
which carries a publicly trusted Let's Encrypt certificate, so both
containerd on the node (pulls) and dockerd in the build pod (pushes) accept
it out of the box.
This used to be a real piece of setup, and the history is worth keeping
because reintroducing a nip.io registry name would bring it all back. A
nip.io address cannot have a public certificate — it is not on the public
suffix list, and every `*.nip.io` certificate on the internet shares one
rate limit — so cert-manager signed Harbor from a private CA instead. The
node pool was told to trust that CA at creation, which covered pulls; the
push came from dockerd inside the build pod, a separate client with its own
trust store, so `dind-pod.yaml` had to mount the CA at
`/etc/docker/certs.d/<registry host>/ca.crt` as well. The failure when that
mount was missing was thoroughly confusing: pushes failed TLS verification
while pulls of the very same image succeeded.