AWS Sagemaker Orchestrator

Orchestrating your pipelines to run on Amazon Sagemaker.

Sagemaker Pipelines is a serverless ML workflow tool running on AWS. It is an easy way to quickly run your code in a production-ready, repeatable cloud orchestrator that requires minimal setup without provisioning and paying for standby compute.

This component is only meant to be used within the context of a remote ZenML deployment scenario. Usage with a local ZenML deployment may lead to unexpected behavior!

When to use it

You should use the Sagemaker orchestrator if:

  • you're already using AWS.

  • you're looking for a proven production-grade orchestrator.

  • you're looking for a UI in which you can track your pipeline runs.

  • you're looking for a managed solution for running your pipelines.

  • you're looking for a serverless solution for running your pipelines.

How it works

The ZenML Sagemaker orchestrator works with Sagemaker Pipelines, which can be used to construct machine learning pipelines. Under the hood, for each ZenML pipeline step, it creates a SageMaker PipelineStep, which contains a Sagemaker Processing job. Currently, other step types are not supported.

How to deploy it

Would you like to skip ahead and deploy a full ZenML cloud stack already, including a Sagemaker orchestrator? Check out the in-browser stack deployment wizard, the stack registration wizard, or the ZenML AWS Terraform module for a shortcut on how to deploy & register this stack component.

In order to use a Sagemaker AI orchestrator, you need to first deploy ZenML to the cloud. It would be recommended to deploy ZenML in the same region as you plan on using for Sagemaker, but it is not necessary to do so. You must ensure that you are connected to the remote ZenML server before using this stack component.

The only other thing necessary to use the ZenML Sagemaker orchestrator is enabling the relevant permissions for your particular role.

How to use it

To use the Sagemaker orchestrator, we need:

  • The ZenML aws and s3 integrations installed. If you haven't done so, run

zenml integration install aws s3
  • Docker installed and running.

  • A remote artifact store as part of your stack (configured with an authentication_secret attribute).

  • A remote container registry as part of your stack.

  • An IAM role or user with an AmazonSageMakerFullAccess managed policy applied to it as well as sagemaker.amazonaws.com added as a Principal Service. Full details on these permissions can be found here or use the ZenML recipe (when available) which will set up the necessary permissions for you.

  • The local client (whoever is running the pipeline) will also have to have the necessary permissions or roles to be able to launch Sagemaker jobs. (This would be covered by the AmazonSageMakerFullAccess policy suggested above.)

There are three ways you can authenticate your orchestrator and link it to the IAM role you have created:

The recommended way to authenticate your SageMaker orchestrator is by registering an AWS Service Connector and connecting it to your SageMaker orchestrator:

zenml service-connector register <CONNECTOR_NAME> --type aws -i
zenml orchestrator register <ORCHESTRATOR_NAME> \
    --flavor=sagemaker \
    --execution_role=<YOUR_IAM_ROLE_ARN>
zenml orchestrator connect <ORCHESTRATOR_NAME> --connector <CONNECTOR_NAME>
zenml stack register <STACK_NAME> -o <ORCHESTRATOR_NAME> ... --set

ZenML will build a Docker image called <CONTAINER_REGISTRY_URI>/zenml:<PIPELINE_NAME> which includes your code and use it to run your pipeline steps in Sagemaker. Check out this page if you want to learn more about how ZenML builds these images and how you can customize them.

You can now run any ZenML pipeline using the Sagemaker orchestrator:

python run.py

If all went well, you should now see the following output:

Steps can take 5-15 minutes to start running when using the Sagemaker Orchestrator.
Your orchestrator 'sagemaker' is running remotely. Note that the pipeline run will only show up on the ZenML dashboard once the first step has started executing on the remote infrastructure.

If it is taking more than 15 minutes for your run to show up, it might be that a setup error occurred in SageMaker before the pipeline could be started. Checkout the Debugging SageMaker Pipelines section for more information on how to debug this.

Sagemaker UI

Sagemaker comes with its own UI that you can use to find further details about your pipeline runs, such as the logs of your steps.

To access the Sagemaker Pipelines UI, you will have to launch Sagemaker Studio via the AWS Sagemaker UI. Make sure that you are launching it from within your desired AWS region.

Once the Studio UI has launched, click on the 'Pipeline' button on the left side. From there you can view the pipelines that have been launched via ZenML:

Debugging SageMaker Pipelines

If your SageMaker pipeline encounters an error before the first ZenML step starts, the ZenML run will not appear in the ZenML dashboard. In such cases, use the SageMaker UI to review the error message and logs. Here's how:

  • Open the corresponding pipeline in the SageMaker UI as shown in the SageMaker UI Section,

  • Open the execution,

  • Click on the failed step in the pipeline graph,

  • Go to the 'Output' tab to see the error message or to 'Logs' to see the logs.

Alternatively, for a more detailed view of log messages during SageMaker pipeline executions, consider using Amazon CloudWatch:

  • Search for 'CloudWatch' in the AWS console search bar.

  • Navigate to 'Logs > Log groups.'

  • Open the '/aws/sagemaker/ProcessingJobs' log group.

  • Here, you can find log streams for each step of your SageMaker pipeline executions.

Run pipelines on a schedule

The ZenML Sagemaker orchestrator doesn't currently support running pipelines on a schedule. We maintain a public roadmap for ZenML, which you can find here. We welcome community contributions (see more here) so if you want to enable scheduling for Sagemaker, please do let us know!

Configuration at pipeline or step level

When running your ZenML pipeline with the Sagemaker orchestrator, the configuration set when configuring the orchestrator as a ZenML component will be used by default. However, it is possible to provide additional configuration at the pipeline or step level. This allows you to run whole pipelines or individual steps with alternative configurations. For example, this allows you to run the training process with a heavier, GPU-enabled instance type, while running other steps with lighter instances.

Additional configuration for the Sagemaker orchestrator can be passed via SagemakerOrchestratorSettings. Here, it is possible to configure processor_args, which is a dictionary of arguments for the Processor. For available arguments, see the Sagemaker documentation . Currently, it is not possible to provide custom configuration for the following attributes:

  • image_uri

  • instance_count

  • sagemaker_session

  • entrypoint

  • base_job_name

  • env

For example, settings can be provided in the following way:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    instance_type="ml.m5.large",
    volume_size_in_gb=30,
)

They can then be applied to a step as follows:

@step(settings={"orchestrator": sagemaker_orchestrator_settings})

For example, if your ZenML component is configured to use ml.c5.xlarge with 400GB additional storage by default, all steps will use it except for the step above, which will use ml.t3.medium (for Processing Steps) or ml.m5.xlarge (for Training Steps) with 30GB additional storage. See the next section for details on how ZenML decides which Sagemaker Step type to use.

Check out this docs page for more information on how to specify settings in general.

For more information and a full list of configurable attributes of the Sagemaker orchestrator, check out the SDK Docs .

Using Warm Pools for your pipelines

Warm Pools in SageMaker can significantly reduce the startup time of your pipeline steps, leading to faster iterations and improved development efficiency. This feature keeps compute instances in a "warm" state, ready to quickly start new jobs.

To enable Warm Pools, use the SagemakerOrchestratorSettings class:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    keep_alive_period_in_seconds = 300, # 5 minutes, default value
)

This configuration keeps instances warm for 5 minutes after each job completes, allowing subsequent jobs to start faster if initiated within this timeframe. The reduced startup time can be particularly beneficial for iterative development processes or frequently run pipelines.

If you prefer not to use Warm Pools, you can explicitly disable them:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    keep_alive_period_in_seconds = None,
)

By default, the SageMaker orchestrator uses Training Steps where possible, which can offer performance benefits and better integration with SageMaker's training capabilities. To disable this behavior:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    use_training_step = False
)

These settings allow you to fine-tune your SageMaker orchestrator configuration, balancing between faster startup times with Warm Pools and more control over resource usage. By optimizing these settings, you can potentially reduce overall pipeline runtime and improve your development workflow efficiency.

S3 data access in ZenML steps

In Sagemaker jobs, it is possible to access data that is located in S3. Similarly, it is possible to write data from a job to a bucket. The ZenML Sagemaker orchestrator supports this via the SagemakerOrchestratorSettings and hence at component, pipeline, and step levels.

Import: S3 -> job

Importing data can be useful when large datasets are available in S3 for training, for which manual copying can be cumbersome. Sagemaker supports File (default) and Pipe mode, with which data is either fully copied before the job starts or piped on the fly. See the Sagemaker documentation referenced above for more information about these modes.

Note that data import and export can be used jointly with processor_args for maximum flexibility.

A simple example of importing data from S3 to the Sagemaker job is as follows:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    input_data_s3_mode="File",
    input_data_s3_uri="s3://some-bucket-name/folder"
)

In this case, data will be available at /opt/ml/processing/input/data within the job.

It is also possible to split your input over channels. This can be useful if the dataset is already split in S3, or maybe even located in different buckets.

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    input_data_s3_mode="File",
    input_data_s3_uri={
        "train": "s3://some-bucket-name/training_data",
        "val": "s3://some-bucket-name/validation_data",
        "test": "s3://some-other-bucket-name/testing_data"
    }
)

Here, the data will be available in /opt/ml/processing/input/data/train, /opt/ml/processing/input/data/val and /opt/ml/processing/input/data/test.

In the case of using Pipe for input_data_s3_mode, a file path specifying the pipe will be available as per the description written here . An example of using this pipe file within a Python script can be found here .

Export: job -> S3

Data from within the job (e.g. produced by the training process, or when preprocessing large data) can be exported as well. The structure is highly similar to that of importing data. Copying data to S3 can be configured with output_data_s3_mode, which supports EndOfJob (default) and Continuous.

In the simple case, data in /opt/ml/processing/output/data will be copied to S3 at the end of a job:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    output_data_s3_mode="EndOfJob",
    output_data_s3_uri="s3://some-results-bucket-name/results"
)

In a more complex case, data in /opt/ml/processing/output/data/metadata and /opt/ml/processing/output/data/checkpoints will be written away continuously:

sagemaker_orchestrator_settings = SagemakerOrchestratorSettings(
    output_data_s3_mode="Continuous",
    output_data_s3_uri={
        "metadata": "s3://some-results-bucket-name/metadata",
        "checkpoints": "s3://some-results-bucket-name/checkpoints"
    }
)

Using multichannel output or output mode except EndOfJob will make it impossible to use TrainingStep and also Warm Pools. See corresponding section of this document for details.

Tagging SageMaker Pipeline Executions and Jobs

The SageMaker orchestrator allows you to add tags to your pipeline executions and individual jobs. Here's how you can apply tags at both the pipeline and step levels:

from zenml import pipeline, step
from zenml.integrations.aws.flavors.sagemaker_orchestrator_flavor import SagemakerOrchestratorSettings

# Define settings for the pipeline
pipeline_settings = SagemakerOrchestratorSettings(
    pipeline_tags={
        "project": "my-ml-project",
        "environment": "production",
    }
)

# Define settings for a specific step
step_settings = SagemakerOrchestratorSettings(
    tags={
        "step": "data-preprocessing",
        "owner": "data-team"
    }
)

@step(settings={"orchestrator": step_settings})
def preprocess_data():
    # Your preprocessing code here
    pass

@pipeline(settings={"orchestrator": pipeline_settings})
def my_training_pipeline():
    preprocess_data()
    # Other steps...

# Run the pipeline
my_training_pipeline()

In this example:

  • The pipeline_tags are applied to the entire SageMaker pipeline object. SageMaker automatically applies the pipeline_tags to all its associated jobs.

  • The tags in step_settings are applied to the specific SageMaker job for the preprocess_data step.

This approach allows for more granular tagging, giving you flexibility in how you categorize and manage your SageMaker resources. You can view and manage these tags in the AWS Management Console, CLI, or API calls related to your SageMaker resources.

Enabling CUDA for GPU-backed hardware

Note that if you wish to use this orchestrator to run steps on a GPU, you will need to follow the instructions on this page to ensure that it works. It requires adding some extra settings customization and is essential to enable CUDA for the GPU to give its full acceleration.

ZenML Scarf

Last updated