Skip to content

014-integration-testing.md

Latest commit

 

History

History
41 lines (25 loc) · 2.59 KB

014-integration-testing.md

File metadata and controls

41 lines (25 loc) · 2.59 KB

ADR-014: E2E integration testing against dedicated k3s cluster

Date: 2026-04-08 Status: Accepted Owner: @tomkis

Context

Schedule CRUD was the first feature requiring API-level tests. The API server is a thin tRPC layer over K8s ConfigMap CRUD — most logic lives in label selectors, YAML serialization, and cross-resource validation. The system has two producers/consumers of schedule ConfigMaps: the API server (writes spec.yaml, reads status.yaml) and the Go controller (reads spec.yaml, writes status.yaml via in-process cron scheduler — see ADR-008).

Decision

Full e2e tests that spin up a dedicated test cluster (mise run cluster:install --vm-name=humr-k3s-test), deploy the complete Humr stack, and test against it.

Test layers

Layer 1: API server CRUD — calls real tRPC endpoints via ingress, verifies responses, error handling, and input validation.

Layer 2: Controller reconciliation (skipped pending #34) — creates a schedule with * * * * * cron, waits for the controller to fire and write status.yaml. Currently blocked by controller informer stalling after OneCLI registration errors.

Infrastructure

  • Dedicated Lima VM (humr-k3s-test) with separate ports (5555 Traefik, 16445 k8s API) to avoid collisions with dev cluster
  • Full Humr chart deployed (API server, controller, UI, OneCLI, PostgreSQL)
  • mise run api-server:e2e orchestrates: build images → create cluster → run vitest → tear down cluster
  • Vitest as test framework, @trpc/client for endpoint calls, @kubernetes/client-node for K8s assertions
  • Cluster parameterization via --vm-name and --lima-template flags on cluster:install/cluster:delete

Alternatives Considered

Ephemeral namespace in existing dev cluster — faster (~3s) but doesn't test the full stack (no controller, no OneCLI). Also pollutes the dev cluster state and couples test runs to dev cluster availability.

Ephemeral namespace in dev cluster with test code deployed — reuses the existing VM but deploys test-specific resources into a separate namespace. Avoids VM creation overhead but shares controller/OneCLI with dev, making tests non-isolated and order-dependent.

Unit tests with mocked contexts — rejected because the interesting logic lives in K8s interactions, not the tRPC router.

Consequences

  • Full test run takes ~3 min (image builds cached, VM creation ~1 min, cert-manager ~30s, helm install ~30s)
  • Test cluster is torn down after each run
  • cluster:install/cluster:delete now accept --vm-name and --lima-template, enabling multiple concurrent clusters