Glossary

Terms that appear in plnt code, docs, and the site — with the definition we mean.

---

Canary

The initial low-traffic-percent rollout of a new model. Default 5%. If the smoke test passes, canary is promoted to 100%; if it fails, helm rollback reverts.

ConfigMap-driven registry

The playground API’s model list is read once at pod startup from a mounted ConfigMap. helm upgrade with a new list restarts the pod and picks up the new registry. There is no REST mutation API.

CRD

Custom Resource Definition — a Kubernetes-native way to add a new “kind” of object. plnt defines one: apiVersion: plnt.work/v1, kind: WorkflowRun.

Deploy saga

The Temporal workflow that turns an WorkflowRun apply into a running, traffic-served pod. Five activities: validate → pull-and-verify weights → helm-install-canary → smoke-test → promote-or-rollback. Compensation on failure. See Deploy saga.

FastAPI

The Python HTTP framework the playground API is built on. Handles routing, pydantic validation, OpenAPI schema generation, and Server-Sent Events.

HPA

Horizontal Pod Autoscaler. Kubernetes primitive for scaling a Deployment on CPU / memory / custom metrics. Every runtime chart ships one.

HTTPBackend

RuntimeAdapter implementation that proxies to any OpenAI-compat upstream over HTTP. vLLM, TGI (--openai-api), SGLang, llama.cpp server all satisfy this contract.

WorkflowRun

The plnt CRD. The declarative unit of “I want this model, served by this runtime, on this cluster.” Applied by kubectl apply -f or plnt deploy. Watched by the plnt operator.

JTBD

Job-to-be-done. Framing from Clay Christensen. What is the user actually trying to accomplish when they reach for plnt?

Kind (cluster)

kind — Kubernetes IN Docker. A local single-node cluster useful for demo and testing. plnt’s local demo path uses kind. See Kind demo.

kopf

Python framework for writing Kubernetes operators. The plnt operator uses it.

MockBackend

RuntimeAdapter implementation that echoes the last user message word-by-word via SSE. No network, no GPU. Used on kind and in tests.

microagents

The workflow-recipe registry that plnt pulls from. Each recipe is a versioned spec (workflow.yaml) + runnable code, stored on S3 or an OCI registry and referenced by name in a WorkflowRun spec. Sibling repo: plnt-work/microagents.

Playground

The public demo surface at plnt.work/playground (UI, owned by plnt-site) and playground.plnt.work (API, owned by plnt).

Registry (plnt-scope)

The playground API’s in-memory model index. Not to be confused with container registry (DOCR, GHCR) or model registry (MLflow).

Runtime

An ML inference server: vLLM, TGI, TRT-LLM, SGLang, mock. Each has a Helm chart in plnt/charts/{runtime}-runtime/.

RuntimeAdapter

Python Protocol in plnt/runtime/base.py that abstracts “how do I talk to a runtime.” Implementations: MockBackend, HTTPBackend. See Runtime adapter.

Saga

Long-running distributed transaction with compensation. The plnt DeployModelWorkflow is a saga: each activity’s failure triggers a helm rollback compensation.

Smoke test

The gate activity in the deploy saga. Runs N test prompts against the canary; asserts TTFT p50 ≤ threshold. Fails the saga if the model doesn’t meet the perf bar.

SSE

Server-Sent Events. Unidirectional stream from server to client over HTTP, Content-Type: text/event-stream. Used by the playground API for streaming chat completions.

Temporal

Durable workflow engine. Owns the deploy saga’s event history and retry state — outlives pod restarts and controller crashes.

TPOT

Time Per Output Token. The steady-state per-token latency during a streamed completion. Reported by plnt bench.

TTFT

Time To First Token. Latency from request accepted to first chunk sent. The primary gate metric in the smoke test.

values.yaml

The Helm chart’s configuration document. Every runtime chart shares a common core (model.name, resources.gpu, replicas.min/max).

vLLM

The reference OSS inference runtime. Paged attention + prefix cache + continuous batching. First runtime plnt ships an end-to-end chart for.