Deployments Overview
Deployments turns a versioned workflow graph into concrete jobs, transports, and assignments on workers. Use it when a system needs multiple stages, branching, durable boundaries, controlled placement, or repeatable plan/apply.
The operator journey
Section titled “The operator journey”- Choose or author: start from the Catalog, a Provider Pack, or a new workflow.
- Design: decide routing, transport, and placement/replicas.
- Validate and publish: resolve graph errors and publish an immutable workflow version.
- Plan: preview generated jobs, channels, placement, warnings, and the difference from current state.
- Apply: commit only the reviewed fresh plan.
- Observe: wait for assignment and readiness; inspect runtime evidence.
- Verify: read back any external destination effect you intend to claim.
- Correct or remove: re-plan a new desired state or tear down only exact run-owned identities.
Pick the next guide by task
Section titled “Pick the next guide by task”| Task | Guide |
|---|---|
| Learn graphs, data edges, messages, and state | Core Concepts |
| Draw, validate, version, and publish a graph | Workflows |
| Clone, distribute, filter, or join streams | Routing and Fan-out |
| Choose local, durable, object, or broker handoff | Transports and Durability |
| Choose worker groups, replicas, and safe ownership | Placement and Scaling |
| Plan, apply, inspect, correct, and tear down | Deployment Manager |
| Automate with an assistant | MCP workflow lifecycle |
Prove it with a tutorial
Section titled “Prove it with a tutorial”- File-store roundtrip: first durable workflow and plan/apply cycle.
- Local fan-out: eight synthetic records, clone routing, and exact branch counts on one worker.
- Application log router: normalization, validation, quarantine, archive, and critical-event routing.
The evidence boundary
Section titled “The evidence boundary”A workflow describes intent. A published version freezes that intent. A plan describes proposed concrete changes. Apply commits desired state. Reconciliation reports convergence. Worker evidence shows observed execution. Only a destination receipt/readback proves an external effect. Keep those statements separate in reviews and incident reports.
For file-based desired-state promotion, continue to GitOps CLI Beta.