tests/gitlab-e2e.sh tests vk-gitlab against a real GitLab CE and fleet on one host: GitLab
in a vk microVM, then vk-hub (TLS, with a CA made for the run), a vk-registry for caches,
one vk node run enrolled in pool e2e, and vk-gitlab run, enrolled as a
relay with vk-gitlab join. The hub approves the GitLab server as e2e and the test project
beforehand. The registry joins the hub with [hub] tokens = true, and the node gives each job
its own registry token (guest_credential = "job");
E2E_TOKENS=0 runs the fleet without tokens instead, the node's own credential carrying the
caches. Each scenario pushes a branch with its own .gitlab-ci.yml and checks the outcome
through GitLab's API.
It is not part of the release gate: tests/release-e2e.sh lists it as not run unless
E2E_GITLAB=1 is set or the script is named on its command line.
./build.sh --fast # dist/vk, vk-hub, vk-registry, vk-gitlab
tests/gitlab-e2e.sh # every scenario
tests/gitlab-e2e.sh --scenario cache,cancel --keep
tests/gitlab-e2e.sh --stop-gitlab # stop a GitLab left up by --keep- Binaries. The node runs
--vk(default$VK, elsedist/vk);vk-hub,vk-registryandvk-gitlabdefault to the ones beside it, and--vk-hub,--vk-registryand--vk-gitlaboverride each. Without avk-registrythecache,registry,pipeline_buildandbuild_cachescenarios are skipped, and withE2E_TOKENS=0the last three. GitLab boots with thevkonPATH, else--vk(--vk-boot). - Scenarios.
basic,scripts(before/after_script),variables(masked, file),exit_codes(allow_failure: exit_codes),artifacts(upload, exclude, dependency download),cache(miss then hit across two pipelines),registry(a job's token throughregistry.vk: its pipeline's namespace takes a push and a lock,imagesrefuses both),pipeline_build(vk-agent buildin one job, the image booted by later jobs of its pipeline by full reference and aspipeline/build/runner; a second build of the name, a symlink in the context or as the Dockerfile, a direct reference into the pipeline namespace and the image from another pipeline refused),build_cache(the hub'sci-buildreplaced by one isolating a branch; avk-agent buildimage built once for two pipelines of another branch and apart for the isolated one, each scope's layers in the registry, each job's token reading and locking its own scope's and refused the other's),services,timeout,cancel,restart(vk-gitlab killed with SIGKILL mid-job, then resumed). All butrestartrun concurrently;restartruns alone after them. - GitLab.
gitlab/gitlab-ce, pinned by digest intests/gitlab/images.env(which says how to bump it), boots with its own/assets/init-containeras the command of a detachedvk rununder vk-agent. Its three volumes are:diskqcow2 files under~/.cache/vk-gitlab-e2e/gitlab-<digest>/(VK_GITLAB_E2E_CACHE), where the instance — root token, project, variables — is kept across runs;--reset-gitlabdiscards it. GitLab listens on the host's own address, since the job VMs' egress refuses loopback, on a port chosen once per instance;GL_HOSToverrides the address. - Timings. About 11 minutes cold (GitLab up and seeded in 5–6, image conversion
included), 9 from the cached instance (GitLab in 3–4: vk rebuilds the boot rootfs, then
gitlab-ctl reconfigureruns), and 5 against a GitLab still up from--keep, which is reused at once.GL_BOOT_TIMEOUT(1500 s) bounds GitLab's boot. - Resources. KVM;
curl,jq,git,openssl,ss,ip,shuf; access to Docker Hub. About 9 GiB of free memory — GitLab's VM 6 GiB (GL_MEM, withGL_CPUS4), each job VM 1 GiB, up to four at once — and 15 GiB of disk for the GitLab image and its state. - Evidence. Everything lands in
target/e2e/<timestamp>/(--run-dir, which must be absent or empty):logs/(GitLab console, hub, node, registry, vk-gitlab at debug level, one log per scenario),jobs/(each job's JSON and trace), and the components' configs. The run prints a result table and exits non-zero when a scenario fails.