Testing

Three tiers, cheapest first:

make test    # unit tests — spec validation, naming, manifests, backend
             # resolution, patch merging. No cluster needed.
make lint    # ruff
make e2e     # the full loop on a disposable k3d cluster

The e2e harness

scripts/e2e.sh is fully self-contained and never touches your default kubeconfig:

1. creates a throwaway k3d cluster;
2. builds the mayfly images (dragonfly, hello, the patched emulator) from the working tree under their published names and imports them — so e2e always tests your local code and the cluster never pulls;
3. mayfly up examples/env.yaml;
4. runs the in-cluster smoke test — S3 round-trip, RDS control-plane + psql, caches, Kafka produce/consume, dynamo, the ALB, dragonfly's /api and /healthz;
5. mayfly up again (idempotency proof), mayfly down, cluster deleted.

CI runs unit (Python 3.9 + 3.13 matrix) then e2e on every push and PR.

Interactive iteration

Keep a long-lived local cluster instead of paying cluster-create each run:

k3d cluster create mayfly-dev --kubeconfig-update-default=false -p "80:80@loadbalancer"
export KUBECONFIG=$(k3d kubeconfig write mayfly-dev)
mayfly up examples/env.yaml
./examples/smoke-test.sh <namespace>

After changing an app image locally, rebuild + k3d image import under the same name/tag and kubectl rollout restart its deployment — same-tag images don't trigger a rollout by themselves.