Run flows on serverless compute
Learn how to use Prefect push work pools to schedule work on serverless infrastructure without having to run a worker.
Push work pools are a special type of work pool. They allow Prefect Cloud to submit flow runs for execution to serverless computing infrastructure without requiring you to run a worker. Push work pools currently support execution in AWS ECS tasks, Azure Container Instances, Google Cloud Run jobs, and Modal.
In this guide you will:
- Create a push work pool that sends work to Amazon Elastic Container Service (AWS ECS), Azure Container Instances (ACI), Google Cloud Run, or Modal
- Deploy a flow to that work pool
- Execute a flow without having to run a worker process to poll for flow runs
You can automatically provision infrastructure and create your push work pool using the prefect work-pool create
CLI command with the --provision-infra
flag.
This approach greatly simplifies the setup process.
First, you will set up automatic infrastructure provisioning for push work pools. Then you will learn how to manually set up your push work pool.
Automatic infrastructure provisioning
With Perfect Cloud you can provision infrastructure for use with an AWS ECS, Google Cloud Run, ACI push work pool. Push work pools in Prefect Cloud simplify the setup and management of the infrastructure necessary to run your flows. However, setting up infrastructure on your cloud provider can still be a time-consuming process. Prefect dramatically simplifies this process by automatically provisioning the necessary infrastructure for you.
We’ll use the prefect work-pool create
CLI command with the --provision-infra
flag to automatically provision your
serverless cloud resources and set up your Prefect workspace to use a new push pool.
Prerequisites
To use automatic infrastructure provisioning, you need:
- the relevant cloud CLI library installed
- to be authenticated with your cloud provider
Install the AWS CLI, authenticate with your AWS account, and set a default region.
If you already have the AWS CLI installed, be sure to update to the latest version.
You will need the following permissions in your authenticated AWS account:
IAM Permissions:
- iam:CreatePolicy
- iam:GetPolicy
- iam:ListPolicies
- iam:CreateUser
- iam:GetUser
- iam:AttachUserPolicy
- iam:CreateRole
- iam:GetRole
- iam:AttachRolePolicy
- iam:ListRoles
- iam:PassRole
Amazon ECS Permissions:
- ecs:CreateCluster
- ecs:DescribeClusters
Amazon EC2 Permissions:
- ec2:CreateVpc
- ec2:DescribeVpcs
- ec2:CreateInternetGateway
- ec2:AttachInternetGateway
- ec2:CreateRouteTable
- ec2:CreateRoute
- ec2:CreateSecurityGroup
- ec2:DescribeSubnets
- ec2:CreateSubnet
- ec2:DescribeAvailabilityZones
- ec2:AuthorizeSecurityGroupIngress
- ec2:AuthorizeSecurityGroupEgress
Amazon ECR Permissions:
- ecr:CreateRepository
- ecr:DescribeRepositories
- ecr:GetAuthorizationToken
If you want to use AWS managed policies, you can use the following:
- AmazonECS_FullAccess
- AmazonEC2FullAccess
- IAMFullAccess
- AmazonEC2ContainerRegistryFullAccess
The above policies give you all the permissions needed, but are more permissive than necessary.
Docker is also required to build and push images to your registry.
Automatically create a new push work pool and provision infrastructure
To create a new push work pool and configure the necessary infrastructure, run this command for your particular cloud provider:
The --provision-infra
flag automatically sets up your default AWS account to execute
flows with ECS tasks.
In your AWS account, this command creates a new IAM user, IAM policy, and
ECS cluster that uses AWS Fargate, VPC, and ECR repository (if they don’t already exist).
In your Prefect workspace, this command creates an
AWSCredentials
block for storing the generated credentials.
Here’s an abbreviated example output from running the command:
Default Docker build namespace
After infrastructure provisioning completes, you will be logged into your new ECR repository and the default Docker build namespace will be set to the URL of the registry.
While the default namespace is set, you do not need to provide the registry URL when building images as part of your deployment process.
To take advantage of this, you can write your deploy scripts like this:
This builds an image with the tag <ecr-registry-url>/my-image:latest
and push it to the registry.
Your image name needs to match the name of the repository created with your work pool. You can create new repositories in the ECR console.
You’re ready to create and schedule deployments that use your new push work pool. Reminder that no worker is required to run flows with a push work pool.
Use existing resources with automatic infrastructure provisioning
If you already have the necessary infrastructure set up, Prefect detects that at work pool creation and the infrastructure provisioning for that resource is skipped.
For example, here’s how prefect work-pool create my-work-pool --provision-infra
looks when existing Azure resources are detected:
Provision infrastructure for an existing push work pool
If you already have a push work pool set up, but haven’t configured the necessary infrastructure, you can use the
provision-infra
sub-command to provision the infrastructure for that work pool.
For example, you can run the following command if you have a work pool named “my-work-pool”.
Prefect creates the necessary infrastructure for the my-work-pool
work pool and provides you with a summary
of the changes:
This command speeds up your infrastructure setup process.
As with the examples above, you need to have the related cloud CLI library installed and to be authenticated with your cloud provider.
Manual infrastructure provisioning
If you prefer to set up your infrastructure manually, exclude the --provision-infra
flag in the CLI command.
In the examples below, you’ll create a push work pool through the Prefect Cloud UI.
To push work to ECS, AWS credentials are required.
Create a user and attach the AmazonECS_FullAccess permissions.
From that user’s page, create credentials and store them somewhere safe for use in the next section.
Work pool configuration
The push work pool stores information about what type of infrastructure the flow will run on, what default values to provide to compute jobs, and other important execution environment parameters. Because the push work pool needs to integrate securely with your serverless infrastructure, you need to store your credentials in Prefect Cloud by making a block.
Create a Credentials block
Navigate to the blocks page, click create new block, and select AWS Credentials for the type.
For use in a push work pool, set the region, access key, and access key secret.
Provide any other optional information and create your block.
Create a push work pool
Now navigate to the work pools page. Click Create to configure your push work pool by selecting a push option in the infrastructure type step.
Each step has several optional fields that are detailed in the work pools documentation. Select the block you created under the AWS Credentials field. This allows Prefect Cloud to securely interact with your ECS cluster.
Create your pool to be ready to deploy flows to your Push work pool.
Deployment
You need to configure your deployment to send flow runs to your push work pool.
For example, if you create a deployment through the interactive command line experience,
choose the work pool you just created. If you are deploying an existing prefect.yaml
file, the deployment would contain:
Deploying your flow to the my-push-pool
work pool ensures that runs that are ready for execution are
submitted immediately—without the need for a worker to poll for them.
Serverless infrastructure may require a certain image architecture
Serverless infrastructure may assume a certain Docker image architecture; for example,
Google Cloud Run will fail to run images built with linux/arm64
architecture. If using Prefect to build your image,
you can change the image architecture through the platform
keyword (for example, platform="linux/amd64"
).
Putting it all together
With your deployment created, navigate to its detail page and create a new flow run. You’ll see the flow start running without polling the work pool, because Prefect Cloud securely connected to your serverless infrastructure, created a job, ran the job, and reported on its execution.
Next steps
Learn more about work pools and workers.
Learn about installing dependencies at runtime or baking them into your Docker image in the Deploy to Docker guide.
Was this page helpful?