Skip to main content

Temporal Nexus - Python SDK feature guide

View Markdown

Use Temporal Nexus to connect Temporal Applications within and across Namespaces using a Nexus Endpoint, a Nexus Service contract, and Nexus Operations.

tip

New to Nexus? Start with the Nexus Python Quickstart.

caution

This Feature Guide includes the new Nexus developer experience: pre-release APIs for the Temporal Operation Handler, Nexus Standalone Activity, and (where supported) the Nexus Code Generator. These APIs are experimental and may change.

This page shows how to do the following:

info

This documentation uses source code derived from the Python Nexus sample.

Run the Temporal Development Server with Nexus enabled​

Prerequisites:

The first step in working with Temporal Nexus involves starting a Temporal Server with Nexus enabled.

temporal server start-dev

This command automatically starts the Temporal development server with the Web UI, and creates the default Namespace. It uses an in-memory database, so do not use it for real use cases.

The Temporal Web UI should now be accessible at http://localhost:8233, and the Temporal Server should now be available for client connections on localhost:7233.

Create caller and handler Namespaces​

Before setting up Nexus endpoints, create separate Namespaces for the caller and handler.

temporal operator namespace create --namespace my-target-namespace
temporal operator namespace create --namespace my-caller-namespace

my-target-namespace will contain the Nexus Operation handler, and we will use a Workflow in my-caller-namespace to call that Operation handler. We use different namespaces to demonstrate cross-Namespace Nexus calls.

Create a Nexus Endpoint to route requests from caller to handler​

After establishing caller and handler Namespaces, the next step is to create a Nexus Endpoint to route requests.

temporal operator nexus endpoint create \
  --name my-nexus-endpoint-name \
  --target-namespace my-target-namespace \
  --target-task-queue my-handler-task-queue

You can also use the Web UI to create the Namespaces and Nexus endpoint.

Define the Nexus Service contract​

Defining a clear contract for the Nexus Service is crucial for smooth communication.

In this example, there is a service module that describes the Service and Operation names along with input/output types for caller Workflows to use the Nexus Endpoint.

You can hand-write that module, but the preferred way is to generate it with the Nexus Code Generator. You write the contract once as a JSON definition file and run nexgen against it, and it emits the typed models, runtime validators, and the Service definition itself.

This is what makes a Nexus Service polyglot. Both sides generate from the same definition file: the handler implements the Service, the caller invokes its Operations, and neither hand-writes a request or response type. A Python handler and a Go caller share no code, but they both run off that same service contract - so they interoperate with no coordination between the teams beyond the contract itself.

The generated validators check every payload against the contract, when a value is parsed off the wire and again when it is serialized onto it, so bad data is rejected at the boundary rather than reaching your Workflow. A value validates identically in every language, which is what lets a caller and a handler written in different languages trust the same contract. See the chat.nexusrpc.yaml sample contract and the Definition files section of the nexgen README for the file format.

Develop a Nexus Service and Operation handlers​

Nexus Operation handlers are typically defined in the same Worker as the underlying Temporal primitives they abstract. Operation handlers can decide if a given Nexus Operation will be synchronous or asynchronous. They can invoke underlying Temporal primitives such as a Query, Signal, or Update using the Temporal SDK Client, or run other reliable code. Use a synchronous Nexus Operation only when its complete execution path is highly reliable, has predictably low latency, and finishes well within the 10-second handler deadline. Use an asynchronous Nexus Operation when latency or availability is uncertain, the work might exceed the handler deadline, or execution depends on a potentially unreliable service or database. Handlers should be reliable since the circuit breaker trips after 5 consecutive retryable errors, blocking all Operations from the caller to that Endpoint.

Every Operation is written with TemporalOperationHandler. The @nexus.temporal_operation decorator hands your start method three things: a context, a Client, and the Operation input. What you do with the Client decides what backs the Operation:

  • Synchronous. Return nexus.TemporalOperationResult.sync(...) and the Operation completes during the handler call. The caller has its result as soon as the call returns.
  • Asynchronous. Call start_workflow, start_activity, or start_workflow_update on the Client. The handler returns as soon as that Execution has started, and the Operation stays open until the Execution finishes, which may be days later. Its result is delivered to the caller through the Nexus completion callback. This is what lets an Operation outlive the Nexus request timeout.

A handler can perform any number of synchronous side effects, such as sending a Signal, but at most one asynchronous backing per invocation.

Develop a Synchronous Nexus Operation handler​

Return a synchronous result when the Operation can answer immediately. The handler computes the answer and returns it, and the Operation completes during the call.

Handlers should be reliable to avoid tripping the circuit breaker, and the whole call has to finish inside the Nexus request timeout.

import nexusrpc
from temporalio import nexus


@nexusrpc.handler.service_handler(service=MyNexusService)
class MyNexusServiceHandler:
    @nexus.temporal_operation
    async def echo(
        self,
        _ctx: nexus.TemporalStartOperationContext,
        client: nexus.TemporalNexusClient,
        input: EchoInput,
    ) -> nexus.TemporalOperationResult[EchoOutput]:
        return nexus.TemporalOperationResult.sync(EchoOutput(message=input.message))

Use the Temporal Client for Signals, Queries, and Updates​

A common pattern is to reach a Workflow that is already running. Query it or Signal it from a synchronous Operation, or use Signal-With-Start to make sure the Workflow exists before the Signal arrives. Those calls complete during the handler call, so they have to finish inside the Nexus request timeout.

Updates are the exception. Do not wait for one inside the handler. Start it with start_workflow_update and it backs the Operation. The handler returns straight away, and the Operation completes when the Update does, however long it takes.

The nexus_messaging sample shows a Nexus Service that Queries and Signals a running Workflow from synchronous Operations, and backs an Operation with a Workflow Update.

The Client your handler receives is not an ordinary Temporal Client. It propagates bidirectional links and request Ids on every call, so the caller-side and handler-side Executions are connected in the UI without wiring anything. Reach the Workflow Client through client.client rather than constructing your own.

In this example the Workflow Id is derived from an identifier carried in the Operation input, so the caller only needs the identifier it cares about:

@nexusrpc.handler.service_handler(service=NexusGreetingService)
class NexusGreetingServiceHandler:
    def _get_workflow_handle(
        self, client: Client, user_id: str
    ) -> WorkflowHandle[GreetingWorkflow, str]:
        return client.get_workflow_handle_for(
            GreetingWorkflow.run, f"GreetingWorkflow_for_{user_id}"
        )

    @nexus.temporal_operation
    async def approve(
        self,
        _ctx: nexus.TemporalStartOperationContext,
        client: nexus.TemporalNexusClient,
        input: ApproveInput,
    ) -> nexus.TemporalOperationResult[ApproveOutput]:
        await self._get_workflow_handle(client.client, input.user_id).signal(
            GreetingWorkflow.approve, input
        )
        return nexus.TemporalOperationResult.sync(ApproveOutput())

There are two examples of messaging through Nexus in the sample code, caller pattern and on-demand pattern. The caller pattern shows how to send messages to an existing Workflow, while the on-demand pattern shows how to start a Workflow through Nexus and then send Signals to it.

Develop an Asynchronous Nexus Operation handler to start a Workflow​

Call start_workflow on the Client. The Operation completes when the Workflow returns, and the Workflow's return value is delivered to the caller as the Operation's result.

@nexusrpc.handler.service_handler(service=MyNexusService)
class MyNexusServiceHandler:
    @nexus.temporal_operation
    async def hello(
        self,
        _ctx: nexus.TemporalStartOperationContext,
        client: nexus.TemporalNexusClient,
        input: HelloInput,
    ) -> nexus.TemporalOperationResult[HelloOutput]:
        return await client.start_workflow(
            HelloHandlerWorkflow.run,
            input,
            id=f"hello-{input.name}-{input.language}",
        )

Workflow IDs should typically be business-meaningful IDs and are used to dedupe Workflow starts. In general, the ID should be passed in the Operation input as part of the Nexus Service contract.

RESOURCES

Attach multiple Nexus callers to a handler Workflow with a Conflict-Policy of Use-Existing.

Map a Nexus Operation input to multiple Workflow arguments​

A Nexus Operation can only take one input parameter. To start a Workflow that takes several, pass them to start_workflow as args instead of a single positional argument:

return await client.start_workflow(
    HelloHandlerWorkflow.run,
    args=[input.name, input.language],
    id=f"hello-{input.name}-{input.language}",
)

Register a Nexus Service in a Worker​

After developing an asynchronous Nexus Operation handler to start a Workflow, the next step is to register your Nexus Service handler in a Worker. At this stage you can pass any arguments you need to your service handler's __init__ method.

hello_nexus/handler/worker.py

async def main():
    client = await Client.connect("localhost:7233", namespace=NAMESPACE)
    worker = Worker(
        client,
        task_queue=TASK_QUEUE,
        workflows=[HelloHandlerWorkflow],
        nexus_service_handlers=[MyNexusServiceHandler()],
    )
    await worker.run()

Develop a caller Workflow that uses the Nexus Service​

To execute a Nexus Operation from the caller Workflow, import the necessary service definition and operation input/output types:

hello_nexus/caller/workflows.py

from temporalio import workflow

with workflow.unsafe.imports_passed_through():
    from hello_nexus.service import MyInput, MyNexusService, MyOutput


@workflow.defn
class CallerWorkflow:
    @workflow.run
    async def run(self, name: str) -> tuple[MyOutput, MyOutput]:
        nexus_client = workflow.create_nexus_client(
            service=MyNexusService,
            endpoint=NEXUS_ENDPOINT,
        )
        # Start the nexus operation and wait for the result in one go, using execute_operation.
        wf_result = await nexus_client.execute_operation(
            MyNexusService.my_workflow_run_operation,
            MyInput(name),
        )
        # Alternatively, you can use start_operation to obtain the operation handle and
        # then `await` the handle to obtain the result.
        sync_operation_handle = await nexus_client.start_operation(
            MyNexusService.my_sync_operation,
            MyInput(name),
        )
        sync_result = await sync_operation_handle
        return sync_result, wf_result

Register the caller Workflow in a Worker and start the caller Workflow​

After developing the caller Workflow, the next step is to register it with a Worker.

Finally, the caller Workflow must be started using client.start_workflow() or client.execute_workflow().

These steps are the same as for any normal Workflow. The Python sample combines them in a single application. See hello_nexus/caller/app.py for reference.

Make Nexus calls across Namespaces with a development Server​

In one terminal, run the Temporal worker in the handler namespace:

uv run handler/worker.py

In another terminal, run the Temporal worker in the caller namespace and start the caller workflow:

uv run caller/app.py

Canceling a Nexus Operation​

To cancel a Nexus Operation from within a Workflow, call handle.cancel() on the operation handle. Only asynchronous operations can be canceled in Nexus, since cancellation is sent using an operation token. The Workflow or other resources backing the operation may choose to ignore the cancellation request. If ignored, the operation may enter a terminal state.

When a Nexus operation is started, the caller can specify different cancellation types that control how the caller reacts to cancellation:

  • ABANDON - Do not request cancellation of the operation.
  • TRY_CANCEL - Initiate a cancellation request and immediately report cancellation to the caller. Note that this type doesn't guarantee that cancellation is delivered to the operation handler if the caller exits before the delivery is done.
  • WAIT_REQUESTED - Request cancellation of the operation and wait for confirmation that the request was received. Doesn't wait for actual cancellation.
  • WAIT_COMPLETED - Wait for operation completion. Operation may or may not complete as cancelled.

The default is WAIT_COMPLETED. Users can set a different option for cancellation_type when starting or executing an operation.

Once the caller Workflow completes, the caller's Nexus Machinery stops attempting to cancel operations that have not yet been canceled, letting them run to completion.

It's okay to leave operations running in some use cases. To ensure cancellations are delivered, wait for all pending operations to deliver their cancellation requests before exiting the Workflow.

See the Nexus cancellation sample for reference.

Make Nexus calls across Namespaces in Temporal Cloud​

This section assumes you are already familiar with how to connect a Worker to Temporal Cloud. The Temporal Cloud CLI is used to create Namespaces and the Nexus Endpoint, and mTLS client certificates will be used to securely connect the caller and handler Workers to their respective Temporal Cloud Namespaces.

Install tcld and generate certificates​

Certificate generation is only available in tcld. To install the latest version of tcld, run the following command (on macOS):

brew install temporalio/brew/tcld

If you don't already have certificates, you can generate them for mTLS Worker authentication using the command below:

tcld gen ca --org $YOUR_ORG_NAME --validity-period 1y --ca-cert ca.pem --ca-key ca.key

These certificates will be valid for one year.

Create caller and handler Namespaces​

Before deploying to Temporal Cloud, ensure that the appropriate Namespaces are created for both the caller and handler. If you already have these Namespaces, you don't need to do this.

temporal cloud login

temporal cloud namespace create \
	--name <your-caller-namespace> \
	--region aws-us-west-2 \
	--ca-certificate-file 'path/to/your/ca.pem' \
	--retention-days 1

temporal cloud namespace create \
	--name <your-target-namespace> \
	--region aws-us-west-2 \
	--ca-certificate-file 'path/to/your/ca.pem' \
	--retention-days 1

Alternatively, you can create Namespaces through the UI: https://cloud.temporal.io/Namespaces.

Create a Nexus Endpoint to route requests from caller to handler​

To create a Nexus Endpoint you must have a Developer account role or higher, and have NamespaceAdmin permission on the --target-namespace.

temporal cloud nexus endpoint create \
  --name <my-nexus-endpoint-name> \
  --target-task-queue my-handler-task-queue \
  --target-namespace <my-target-namespace.account> \
  --allow-namespace <my-caller-namespace.account> \
  --description-file hello_nexus/endpoint_description.md

The --allow-namespace is used to build an Endpoint allowlist of caller Namespaces that can use the Nexus Endpoint, as described in Runtime Access Control.

Alternatively, you can create a Nexus Endpoint through the UI: https://cloud.temporal.io/nexus.

Observability​

Web UI​

A synchronous Nexus Operation will surface in the caller Workflow as follows, with just NexusOperationScheduled and NexusOperationCompleted events in the caller's Event history:

Observability Sync

Observability Sync

An asynchronous Nexus Operation will surface in the caller Workflow as follows, with NexusOperationScheduled, NexusOperationStarted, and NexusOperationCompleted, in the caller's Event history:

Observability Async

Observability Async

Temporal CLI​

Use the workflow describe command to show pending Nexus Operations in the caller Workflow and any attached callbacks on the handler Workflow:

temporal workflow describe -w <ID>

Nexus events are included in the caller's Event history:

temporal workflow show -w <ID>

For asynchronous Nexus Operations the following are reported in the caller's history:

  • NexusOperationScheduled
  • NexusOperationStarted
  • NexusOperationCompleted

For synchronous Nexus Operations the following are reported in the caller's history:

  • NexusOperationScheduled
  • NexusOperationCompleted
info

NexusOperationStarted isn't reported in the caller's history for synchronous operations.

Learn more​