Kind demo

Deploys the plnt operator + Temporal + playground API on a local kind cluster, then applies an WorkflowRun and shows the deploy saga running end-to-end.

The saga uses a CPU-stub image in place of vLLM so the flow works without a real GPU. Frame this honestly in a demo: “the platform layer is what you’re seeing; hardening for GPU workloads is where the next phase goes.”

Prereqs (one-time)

brew install kind kubectl helm

The demo — copy/paste in order

1. Cluster + operator install (~90 s)

# create a local single-node cluster
kind create cluster --name plnt

# install the CRD + operator + Temporal
kubectl apply -f plnt/operators/crds/workflowrun.yaml
kubectl create namespace plnt-system
helm install temporal temporal/temporal --namespace plnt-system \
  --set server.replicaCount=1 --set cassandra.config.cluster_size=1

# install the playground API gateway (mock backend, works without any real model)
helm install plnt-playground plnt/charts/playground-api --namespace plnt-system

# start the operator locally (dev mode — in prod this ships as a Deployment)
python -m kopf run plnt/operators/workflowrun_controller.py --namespace=default &

# start the Temporal worker
python -m plnt.workflows.worker &

2. The one command that matters (~30 s)

kubectl apply -f examples/llama-3.1-70b-instruct.yaml

Then in three side-by-side terminal panes:

# pane A — the resource going through phases
kubectl get workflowrun -w

NAME                        RUNTIME  GPU              PHASE           ENDPOINT
llama-3.1-70b-instruct      vllm     nvidia.com/h100  Validating
llama-3.1-70b-instruct      vllm     nvidia.com/h100  PullingWeights
llama-3.1-70b-instruct      vllm     nvidia.com/h100  Deploying
llama-3.1-70b-instruct      vllm     nvidia.com/h100  SmokeTesting
llama-3.1-70b-instruct      vllm     nvidia.com/h100  Promoting
llama-3.1-70b-instruct      vllm     nvidia.com/h100  Ready           llama-3.1-70b-instruct.default.svc:8000

# pane B — the operator picking up the event
kubectl logs -f deployment/plnt-operator -n plnt-system

WorkflowRun/llama-3.1-70b-instruct created — starting DeployModelWorkflow

# pane C — the Temporal workflow history
open http://localhost:7233        # Temporal UI, shows the saga step-by-step

3. Actually call the model (~15 s)

kubectl port-forward svc/plnt-playground 8080:8080 -n plnt-system &

curl -s http://localhost:8080/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"plnt-mock-7b","messages":[{"role":"user","content":"say hi"}]}' \
  | jq .

Response comes back OpenAI-shape. In production the same request hits vLLM / TGI / TRT-LLM / SGLang on real GPUs.

4. Rollback story (~15 s)

kubectl delete workflowrun llama-3.1-70b-instruct

The operator cancels the workflow → the workflow triggers helm uninstall via compensation → kubectl get pods shows the release being removed.

Recording tips

• Use asciinema rec kind-demo.cast — text terminal recording, tiny file, embeds anywhere.
• Or QuickTime → New Screen Recording, cropped to iTerm window.
• Speed up 2× in post — nobody wants to watch the helm install for real.
• One take. Rehearse once; imperfect > polished.

Talking points

1. “This is the deploy contract.” Show examples/llama-3.1-70b-instruct.yaml.
2. “This is the operator watching it.” Flash plnt/operators/workflowrun_controller.py.
3. “This is the saga that does the work.” Flash plnt/workflows/deploy_model.py, point at the 5-step try/except with compensation.
4. “This is the runtime chart.” Show plnt/charts/vllm-runtime/templates/deployment.yaml, highlight the nvidia.com/gpu resource request.
5. “Playground is at play.plnt.work.” Open it, chat with a real model.

Total demo run-time: 3–4 minutes. Rehearsed cold: 6 minutes.