Managed Compute in Python

This feature is currently in beta and may be subject to change.

Overview

Hatchet Managed Compute lets you define and manage compute resources directly in your Python code. This guide shows you how to configure compute instances, create workflows, and manage workers using the Hatchet Python SDK.

This guide assumes you are already familiar with the basics of Hatchet and have a local workflow running using Docker. If you are not in this state, please see the Getting Started Guide and Docker Guide.

Basic Configuration

Compute Configuration

The Compute class allows you to define resource requirements for your workload directly in your Python code. You can define multiple compute configurations to use in your workflows on a step-by-step basis. This allows you to easily optimize your compute resources for different parts of your workflow.

from hatchet_sdk.compute import Compute
 
# Define a default compute configuration
default_compute = Compute(
    cpu_kind="shared",    # CPU type: "shared" or "performance"
    cpus=2,              # Number of CPU cores
    memory_mb=1024,      # Memory allocation in MB
    num_replicas=2,      # Number of instances
    regions=["ewr"]      # Deployment regions
)
 
# Define a basic compute configuration
basic = Compute(
    cpu_kind="shared",
    cpus=1,
    memory_mb=1024,
    num_replicas=1,
    regions=["ewr"]
)

For a full list of configuration options, see the Compute API documentation.

GPU Support

⚠️

GPU compute has limited region support and constraints. See the GPU docs for more information.

Hatchet Managed Compute supports GPU instances. You can define GPU compute configurations in the same way as CPU configurations, but with the addition of the gpu parameter.

gpu_compute = Compute(
    cpu_kind="shared",
    gpu_kind="a100",
    gpus=1,
    cpus=1,
    memory_mb=1024,
    num_replicas=1,
    regions=["ewr"],
)

For a full list of configuration options, see the Compute API documentation.

Defining Compute Requirements in Workflows

Workflow Definition

Use compute configuration can then be passed to the step decorator to define compute requirements for each step in your workflow.

from hatchet_sdk import Context, Hatchet
 
hatchet = Hatchet()
 
@hatchet.workflow(on_events=["user:create"])
class ManagedWorkflow:
    @hatchet.step(
        timeout="11s",
        retries=3,
        compute=default_compute
    )
    def step1(self, context: Context):
        print("executed step1")
        time.sleep(10)
        return {
            "step1": "step1",
        }
 
    @hatchet.step(
        timeout="11s",
        retries=3,
        compute=basic
    )
    def step2(self, context: Context):
        print("executed step2")
        time.sleep(10)
        return {
            "step2": "step2",
        }

Worker Management

Setting Up Workers

Configure and start workers to execute your workflows as you would normally do with any other Hatchet worker:

def main():
    # Create workflow instance
    workflow = ManagedWorkflow()
 
    # Initialize worker with max runs limit
    worker = hatchet.worker("test-worker", max_runs=1)
 
    # Register workflow with worker
    worker.register_workflow(workflow)
 
    # Start the worker
    worker.start()

A complete example of a workflow that uses managed compute can be found here (opens in a new tab).

Docker Self-Hosted