Workflows

Workflows are a sequence of steps that gets executed durably. A workflow can be seen as a special type of Virtual Object with some special characteristics:

• Each workflow definition has a run handler that is annotated with @service_name.main() and implements the workflow logic.
• The run handler executes exactly one time for each workflow instance (object / key).
• A workflow definition can implement other handlers that can be called multiple times, and can interact with the workflow.
• Workflows have access to the WorkflowContext and WorkflowSharedContext, giving them some extra functionality, for example Durable Promises to signal workflows.

Workflow retention time

The retention time of a workflow execution is 24 hours after the finishing of the run handler. After this timeout any K/V state is cleared, the workflow's shared handlers cannot be called anymore, and the Durable Promises are discarded. The retention time can be configured via the Admin API per Workflow definition by setting workflow_completion_retention.

Implementing workflows

Have a look at the code example to get a better understanding of how workflows are implemented:

The run handler

Every workflow needs a run handler. This handler has access to the same SDK features as Service and Virtual Object handlers. For example, use ctx.run to log intermediate results in Restate and avoid re-execution on replay.

Querying workflows

Similar to Virtual Objects, you can retrieve the K/V state of workflows via the other handlers defined in the workflow definition, For example, here we expose the status of the workflow to external clients. Every workflow execution can be seen as a new object, so the state is isolated to a single workflow execution. The state can only be mutated by the run handler of the workflow. The other handlers can only read the state.

Signaling workflows

You can use Durable Promises to interact with your running workflows: to let the workflow block until an event occurs, or to send a signal / information into or out of a running workflow. These promises are durable and distributed, meaning they survive crashes and can be resolved or rejected by any handler in the workflow.

Do the following:

1. Create a promise that is durable and distributed in the run handler.
2. Resolve or reject the promise in another handler in the workflow. This can be done at most one time.

You can also use this pattern in reverse and let the run handler resolve promises that other handlers are waiting on. For example, the run handler could resolve a promise when it finishes a step of the workflow, so that other handlers can request whether this step has been completed.

Serving and registering workflows

You serve workflows in the same way as Services and Virtual Objects: by binding them to an HTTP endpoint. Make sure you register the endpoint in Restate before invoking it.

signup_workflow = Workflow("SignupWorkflow")
email_client = EmailClient()


@signup_workflow.main()
async def run(ctx: WorkflowContext, email: str):
    secret = await ctx.run("secret", lambda: str(uuid.uuid4()))
    ctx.set("onboarding_status", "Generated secret")

    await ctx.run(
        "send email",
        lambda: email_client.send_email_with_link(email, secret)
    )
    ctx.set("onboarding_status", "Sent email")

    click_secret = await ctx.promise("email.clicked").value()
    ctx.set("onboarding_status", "Clicked email")

    return click_secret == secret


@signup_workflow.handler()
async def click(ctx: WorkflowSharedContext, secret: str):
    await ctx.promise("email.clicked").resolve(secret)


@signup_workflow.handler()
async def get_status(ctx: WorkflowSharedContext):
    return await ctx.get("onboarding_status")

app = restate.app([signup_workflow])

Workflows-as-code with Restate

Check out some examples of workflows-as-code with Restate on the use case page.

Submitting workflows from a Restate service

Submit/query/signal: Call the workflow handlers in the same way as for Services and Virtual Objects. This returns the result of the workflow/handler once it has finished. Use ctx.workflow_send to call the handler without waiting for the result. You can only call the run handler (submit) once per workflow ID (here "someone").

@user_management_object.handler()
async def signup_user(ctx: ObjectContext, email: str):
    result = await ctx.workflow_call(run, key="someone", arg=email)


@user_management_object.handler()
async def query_status(ctx: ObjectContext):
    status = await ctx.workflow_call(get_status, key="someone", arg=None)

Submitting workflows over HTTP

Submit/query/signal: Call any handler of the workflow in the same way as for Services and Virtual Objects. This returns the result of the handler once it has finished. Add /send to the path for one-way calls. You can only call the run handler once per workflow ID (here "someone").

curl localhost:8080/signup/someone/run \
    -H 'content-type: application/json' \
    -d '{"email": "someone@restate.dev"}'

Attach/peek: This lets you retrieve the result of a workflow or check if it's finished.

curl localhost:8080/restate/workflow/signup/someone/attach
curl localhost:8080/restate/workflow/signup/someone/output

Inspecting workflows

Have a look at the introspection docs on how to inspect workflows. You can use this to for example:

• Inspect the progress of a workflow by looking at the invocation journal
• Inspect the K/V state of a workflow