fix(e2e): unblock MicroShift SCC diagnostics + bump bootstrap timeout#466
Merged
Merged
Conversation
PR runs against the MicroShift / OpenShift SCC admission lane were
failing with no observable signal: the Eventually at
e2e_test.go:2092 timed out after 120s waiting for the InferenceService
Deployment to appear, and the failure dump showed `kubectl logs
<empty> -n llmkube-system` because controllerPodName was never
populated for focus runs. So we had a flake with no diagnostics.
Two fixes:
1. AfterEach now resolves the controller pod lazily by label selector
when the variable is empty. The Manager Context's BeforeAll is the
only path that previously populated it; -ginkgo.focus on the
OpenShift SCC Context skipped that path entirely.
2. The OpenShift SCC test now:
- Splits the single 120s Eventually into Model→Ready and
Deployment→exists steps so a future failure tells us which
stage stalled (HTTP-source Model goes Ready quickly; the
Deployment then waits on ensureModelCachePVC + reconcileDeployment).
- Bumps the Eventually budget to 5 minutes, matching the
helm install --wait timeout the workflow uses elsewhere.
MicroShift on MINC has materially slower first-reconcile than
kind; 120s was always tight.
- Captures per-test namespace state on failure (Model /
InferenceService / PVC / Pod listing + describe + events), which
the outer AfterEach can't see because it only looks at
llmkube-system.
This does not change any controller behavior or production code path;
it changes the test's diagnostic posture so the next failure (if any)
ships with actionable evidence instead of an empty log line.
Signed-off-by: Christopher Maher <chris@mahercode.io>
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Merged
Defilan
added a commit
to Defilan/LLMKube
that referenced
this pull request
May 14, 2026
The MicroShift / MINC e2e lane has been flaking on the OpenShift SCC admission test with a confusing error that PR defilantech#466's improved diagnostics finally surfaced: deployments.apps "scc-test-inference" is forbidden: cannot set blockOwnerDeletion in this case because cannot find RESTMapping for APIVersion inference.llmkube.dev/v1alpha1 Kind InferenceService: no matches for kind "InferenceService" in version "inference.llmkube.dev/v1alpha1" Trace: controller-runtime's controllerutil.SetControllerReference sets BlockOwnerDeletion=true by default. The kube-apiserver GarbageCollector admission plugin validates that flag by RESTMapping the owner Kind to check the caller's `update` permission on the owner's `finalizers` subresource. On kind the API server's discovery cache is warm by the time the controller starts; on MicroShift-in- MINC, the in-container apiserver populates discovery lazily on first request, and the controller's first reconcile races it and loses. We do not need BlockOwnerDeletion. Cascading delete still cleans up operator-managed Deployments/Services/ConfigMaps/HPAs when their parent ModelRouter or InferenceService is deleted; the "block" semantics only matter for finalizer-based cleanup workflows LLMKube does not use. Trade: cleaner bootstrap for slightly looser cascade ordering guarantees we weren't relying on. Fix: new setControllerReferenceUnblocked helper that wraps controllerutil.SetControllerReference and clears BlockOwnerDeletion on the owner ref it just installed. Applied to all 9 call sites in the controller package (router deployment / service / configmap, both InferenceService and ModelRouter reconcilers, HPA reconciler). Coverage: TestSetControllerReferenceUnblocked verifies the resulting owner ref has Controller=true and BlockOwnerDeletion=false. Existing tests in router_builders_test.go and modelrouter_controller_test.go continue to pass — the change is in the metadata, not the spec. Side benefit: removes the controllerutil import from five files that no longer need it. Signed-off-by: Christopher Maher <chris@mahercode.io>
6 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The OpenShift SCC admission e2e Context has been flaking on multiple PRs (most recently #464) with no diagnostic signal: it times out at
e2e_test.go:2092waiting for the InferenceService Deployment to be created, but the failure dump showskubectl logs -n llmkube-system(empty pod name) instead of controller logs.This PR:
AfterEachby label selector whencontrollerPodNameis empty (so focused CI runs still produce logs).Model→ReadyandDeployment→existscheckpoints so the next failure tells us which step stalled.helm install --waitbudget; MicroShift-on-MINC is materially slower than kind).describe+ events). The outer AfterEach only seesllmkube-system, so the namespace where the SCC story actually plays out was invisible.Why
The MicroShift / MINC lane was failing closed: red CI signal with no actionable evidence. Either we ship without OpenShift coverage at all, or we improve the diagnostic posture so the next failure is debuggable. This is the latter.
The job is still
continue-on-error: truewhile it stabilizes, so this is not a behavior change. It is an observability fix. No controller / production code is touched.Fixes ongoing MicroShift e2e flake observed across recent PRs (#464 most recently).
How
controllerPodNameresolution: when empty, dokubectl get pods -n llmkube-system -l control-plane=controller-manager -o jsonpath={.items[0].metadata.name}and use the result. Adds ~one kubectl call to the failure path only.deferinside the It block runs after the Eventually fails but before the spec returns. Dumpsget/describeon Model, InferenceService, PVC, pods, and events in the test namespace.openshiftEventuallyTimeout = 5 * time.Minuteso the three Eventuallys in this Context share one knob. Was 2 minutes each, now 5 minutes each.Model→Readycheck: HTTP-source Models reach Ready quickly (the controller'sreconcileRuntimeResolvedSourcemarks Ready immediately and defers actual fetch to the InferenceService init container). If this Eventually fails, the bug is in Model reconcile; if Model reaches Ready but Deployment doesn't appear, the bug is inensureModelCachePVCorreconcileDeployment.Checklist
make testpasses locally (e2e package builds with-tags=e2e; unit-test path not exercised by this change)make lintpasses locallygit commit -s) per DCO