# Serverless Workers on GCP Cloud Run - Python SDK

> Run a Temporal Worker on a GCP Cloud Run worker pool using the Python SDK.

> **Pre-release**
> Cloud Run support is in Pre-release, and its APIs may change in backwards-incompatible ways.
> Create a [support ticket](/cloud/support#support-ticket) or contact your account team for access, and
> [sign up for updates](https://temporal.io/pages/serverless-workers-updates) to hear when Cloud Run reaches Public Preview.

On a [GCP Cloud Run worker pool](https://cloud.google.com/run/docs/resource-model#worker-pools), you run a standard long-lived Temporal Worker.
Register Workflows and Activities the same way you would with any other Python Worker, and Temporal Cloud scales the pool up and down as work arrives and drains.

A Cloud Run Worker needs no Cloud Run-specific package.
The one addition to a standard Worker is Worker Versioning, which is required for Serverless Workers.
The `temporalio.contrib.gcp` package adds an OpenTelemetry plugin with defaults suited to a Cloud Run worker pool, covered below.

For the end-to-end deployment guide covering the Worker Pool, IAM, and compute configuration, see [Deploy a Serverless Worker on GCP Cloud Run](/production-deployment/worker-deployments/serverless-workers/cloud-run).

## Create a versioned Worker 

Build the Worker as you would any long-running Python Worker, then pass `deployment_config` to `Worker()` to declare the Worker Deployment Version and turn versioning on.

The following Worker reads its connection settings and Task Queue from the environment, so the same image can run against any Namespace:

```python
import asyncio
import os

from temporalio.client import Client
from temporalio.common import VersioningBehavior, WorkerDeploymentVersion
from temporalio.envconfig import ClientConfig
from temporalio.worker import Worker, WorkerDeploymentConfig

from my_activities import my_activity
from my_workflows import MyWorkflow

async def main() -> None:
    client = await Client.connect(**ClientConfig.load_client_connect_config())

    worker = Worker(
        client,
        task_queue=os.environ["TEMPORAL_TASK_QUEUE"],
        workflows=[MyWorkflow],
        activities=[my_activity],
        deployment_config=WorkerDeploymentConfig(
            version=WorkerDeploymentVersion(
                deployment_name="my-app",
                build_id="build-1",
            ),
            use_worker_versioning=True,
            default_versioning_behavior=VersioningBehavior.PINNED,
        ),
    )
    await worker.run()

if __name__ == "__main__":
    asyncio.run(main())
```

`deployment_name` and `build_id` together identify the Worker Deployment Version. Both values must match the version you create with `temporal worker deployment create-version` in the deployment guide, or the Worker polls under a version the WCI does not manage.

Every Workflow needs a [versioning behavior](/worker-versioning#versioning-behaviors), either `PINNED` or `AUTO_UPGRADE`.
Setting `default_versioning_behavior` as shown above covers every Workflow on the Worker.
To set the behavior per Workflow instead, pass `versioning_behavior` to the `@workflow.defn` decorator:

```python
from temporalio import workflow
from temporalio.common import VersioningBehavior

@workflow.defn(versioning_behavior=VersioningBehavior.PINNED)
class MyWorkflow:
    @workflow.run
    async def run(self, name: str) -> str:
        ...
```

For general Worker setup and options that are not specific to Cloud Run, see [Run a Worker](/develop/python/workers/run-worker-process).

## Configure the Temporal connection 

The `temporalio.envconfig` package loads Temporal Client configuration from environment variables and an optional TOML config file, so the Worker code carries no Namespace or credentials.
Set the non-secret values as environment variables on the Worker Pool, and mount the Temporal Cloud API key or TLS material from Secret Manager.
For the full list of supported variables, the config file format, and profiles, see [Environment configuration](/develop/environment-configuration).

`ClientConfig.load_client_connect_config()` returns the keyword arguments for `Client.connect`, which is why the Worker above unpacks it with `**`.
To inspect or change values before connecting, load the profile instead and convert it yourself:

```python
from temporalio.envconfig import ClientConfigProfile

profile = ClientConfigProfile.load()
connect_config = profile.to_client_connect_config()
client = await Client.connect(**connect_config)
```

## Keep Activities safe across scale-in 

The WCI decides when to remove an instance from Task Queue activity, not from what an individual instance is doing.
An instance running a long Activity can be stopped mid-execution.

Use [Activity Heartbeats](/develop/python/activities/timeouts#activity-heartbeats) so a retry resumes from the last recorded progress instead of starting over:

```python
from temporalio import activity

@activity.defn
async def my_activity(items: list[str]) -> str:
    for i, item in enumerate(items):
        activity.heartbeat(i)
        # ... process item
    return "done"
```

For how scale-in decisions are made, see [Serverless Workers on GCP Cloud Run](/serverless-workers/cloud-run#lifecycle).

## Add observability with OpenTelemetry 

The `temporalio.contrib.gcp.OpenTelemetryPlugin` configures observability with defaults suited to a Cloud Run worker pool.
By default, the plugin configures:

- OTLP gRPC export to `localhost:4317`, the collector sidecar endpoint.
- `service.name` from the Cloud Run-provided `CLOUD_RUN_WORKER_POOL` environment variable.
- A replay-safe OpenTelemetry tracer provider.
- Temporal Core metrics with a 60-second export interval.

The underlying metrics and traces are the same ones the Python SDK emits in any environment.
For general observability concepts and the full list of available metrics, see [Observability - Python SDK](/develop/python/platform/observability) and the [SDK metrics reference](/references/sdk-metrics).

Pass the plugin to `Client.connect`. Opt into `add_temporal_spans=True` to trace named operations such as `RunWorkflow:GreetingWorkflow`:

<!--SNIPSTART python-cloud-run-otel-worker-->
[gcp_open_telemetry/worker.py](https://github.com/temporalio/samples-python/blob/main/gcp_open_telemetry/worker.py)
```py
# Endpoint, service name, Core metrics, and tracer provider all use the GCP
# plugin defaults. The opt-in adds named Temporal operation spans.
plugin = OpenTelemetryPlugin(add_temporal_spans=True)
client = await Client.connect(
    settings.address,
    namespace=settings.namespace,
    api_key=settings.api_key,
    tls=True,
    plugins=[plugin],
)
worker = Worker(
    client,
    task_queue=settings.task_queue,
    workflows=[GreetingWorkflow],
    activities=[compose_greeting],
    graceful_shutdown_timeout=WORKER_GRACEFUL_SHUTDOWN_TIMEOUT,
)
```
<!--SNIPEND-->

Named Temporal operation spans are opt-in. The endpoint, service name, tracer provider, and Core metrics use the plugin defaults, but you must set `add_temporal_spans=True` to emit spans for Workflow and Activity operations.

### Run the collector as a sidecar 

The plugin exports to a collector you run as a second container in the Worker Pool.
Use the [Google-Built OpenTelemetry Collector](https://cloud.google.com/stackdriver/docs/instrumentation/google-built-otel), which detects the Cloud Run resource, authenticates through the Worker Pool's service account, exports metrics to Google Managed Service for Prometheus, and exports traces through the Google Cloud Telemetry API.

Provide the following collector configuration. It receives OTLP on `localhost:4317`, detects the GCP resource, and routes metrics and traces to their respective pipelines:

<!--SNIPSTART python-cloud-run-otel-collector-config-->
[gcp_open_telemetry/collector-config.yaml](https://github.com/temporalio/samples-python/blob/main/gcp_open_telemetry/collector-config.yaml)
```yaml
receivers:
  otlp:
    protocols:
      grpc:
        endpoint: localhost:4317

processors:
  batch/traces:
    send_batch_max_size: 200
    send_batch_size: 200
    timeout: 5s
  memory_limiter:
    check_interval: 1s
    limit_percentage: 65
    spike_limit_percentage: 20
  resource_detection:
    detectors: [gcp]
    timeout: 10s
  transform/collision:
    metric_statements:
      - context: datapoint
        statements:
          - set(attributes["exported_location"], attributes["location"])
          - delete_key(attributes, "location")
          - set(attributes["exported_cluster"], attributes["cluster"])
          - delete_key(attributes, "cluster")
          - set(attributes["exported_namespace"], attributes["namespace"])
          - delete_key(attributes, "namespace")
          - set(attributes["exported_job"], attributes["job"])
          - delete_key(attributes, "job")
          - set(attributes["exported_instance"], attributes["instance"])
          - delete_key(attributes, "instance")
          - set(attributes["exported_project_id"], attributes["project_id"])
          - delete_key(attributes, "project_id")
  transform/set_project_id:
    error_mode: ignore
    trace_statements:
      - set(resource.attributes["gcp.project_id"], resource.attributes["gcp.project.id"]) where resource.attributes["gcp.project.id"] != nil
      - set(resource.attributes["gcp.project_id"], resource.attributes["cloud.account.id"]) where resource.attributes["gcp.project_id"] == nil and resource.attributes["cloud.account.id"] != nil

exporters:
  googlemanagedprometheus:
  otlp_grpc:
    endpoint: telemetry.googleapis.com:443
    compression: none
    balancer_name: pick_first
    auth:
      authenticator: googleclientauth

extensions:
  googleclientauth:
  health_check:
    endpoint: 0.0.0.0:13133

service:
  extensions: [googleclientauth, health_check]
  pipelines:
    metrics:
      receivers: [otlp]
      processors: [memory_limiter, resource_detection, transform/collision]
      exporters: [googlemanagedprometheus]
    traces:
      receivers: [otlp]
      processors:
        [memory_limiter, resource_detection, transform/set_project_id, batch/traces]
      exporters: [otlp_grpc]
```
<!--SNIPEND-->

The metrics pipeline forwards each SDK export directly, without a batch processor, so a runtime shutdown-time export cannot collide with a periodic export on the same Managed Prometheus time series.
Traces use a dedicated five-second batch processor.

To verify telemetry after deploying, confirm that Trace Explorer contains `RunWorkflow:GreetingWorkflow` with `service.name` equal to the Worker Pool name, and that Metrics Explorer contains `prometheus.googleapis.com/temporal_workflow_completed_total/counter`.