Develop a Custom Orchestrator

How to develop a custom orchestrator

Base Implementation

ZenML aims to enable orchestration with any orchestration tool. This is where the BaseOrchestrator comes into play. It abstracts away many of the ZenML specific details from the actual implementation and exposes a simplified interface:

from abc import ABC, abstractmethod
from typing import Any, Type

from zenml.config.pipeline_deployment import PipelineDeployment
from zenml.enums import StackComponentType
from zenml.stack import StackComponent, StackComponentConfig, Stack, Flavor

class BaseOrchestratorConfig(StackComponentConfig):
    """Base class for all ZenML orchestrator configurations."""


class BaseOrchestrator(StackComponent, ABC):
    """Base class for all ZenML orchestrators"""

    @abstractmethod
    def prepare_or_run_pipeline(
        self,
        deployment: PipelineDeployment,
        stack: Stack,
    ) -> Any:
        """Prepares and runs the pipeline outright or returns an intermediate
        pipeline representation that gets deployed.
        """

class BaseOrchestratorFlavor(Flavor):
    """Base orchestrator for all ZenML orchestrator flavors."""

    @property
    @abstractmethod
    def name(self):
        """Returns the name of the flavor."""

    @property
    def type(self) -> StackComponentType:
        """Returns the flavor type."""
        return StackComponentType.ORCHESTRATOR

    @property
    def config_class(self) -> Type[BaseOrchestratorConfig]:
        """Config class for the base orchestrator flavor."""
        return BaseOrchestratorConfig

    @property
    @abstractmethod
    def implementation_class(self) -> Type["BaseOrchestrator"]:
        """Implementation class for this flavor."""

This is a slimmed-down version of the base implementation which aims to highlight the abstraction layer. In order to see the full implementation and get the complete docstrings, please check the source code on GitHub.

Build your own custom orchestrator

If you want to create your own custom flavor for an orchestrator, you can follow the following steps:

1. Create a class which inherits from the BaseOrchestrator class and implement the abstract prepare_or_run_pipeline method.
2. If you need to provide any configuration, create a class which inherits from the BaseOrchestratorConfig class add your configuration parameters.
3. Bring both of the implementation and the configuration together by inheriting from the BaseOrchestratorFlavor class. Make sure that you give a name to the flavor through its abstract property.

Once you are done with the implementation, you can register it through the CLI as:

zenml orchestrator flavor register

It is important to draw attention to when and how these base abstractions are coming into play in a ZenML workflow.

• The CustomOrchestratorFlavor class is imported and utilized upon the creation of the custom flavor through the CLI.
• The CustomOrchestratorConfig class is imported when someone tries to register/update a stack component with this custom flavor. Especially, during the registration process of the stack component, the config will be used to validate the values given by the user. As Config object are inherently pydantic objects, you can also add your own custom validators here.
• The CustomOrchestrator only comes into play when the component is ultimately in use.

The design behind this interaction lets us separate the configuration of the flavor from its implementation. This way we can register flavors and components even when the major dependencies behind their implementation are not installed in our local setting (assuming the CustomOrchestratorFlavor and the CustomOrchestratorConfig are implemented in a different module/path than the actual CustomOrchestrator).

Some additional implementation details

Not all orchestrators are created equal. Here is a few basic categories that differentiate them.

Direct Orchestration

The implementation of a local orchestrator can be summarized as follows: code:

for step in deployment.steps.values():
    self.run_step(step)

The orchestrator basically iterates through each step and directly executes the step within the same Python process. Obviously all kind of additional configuration could be added around this.

Python Operator based Orchestration

The airflow orchestrator has a slightly more complex implementation of the prepare_or_run_pipeline() method. Instead of immediately executing a step, a PythonOperator is created which contains a _step_callable. This _step_callable will ultimately execute the self.run_step(...) method of the orchestrator. The PythonOperators are assembled into an AirflowDag which is returned. Through some Airflow magic, this DAG is loaded by the connected instance of Airflow and orchestration of this DAG is performed either directly or on a set schedule.

Container-based Orchestration

The kubeflow orchestrator is a great example of container-based orchestration. In an implementation-specific method called prepare_pipeline_deployment() a Docker image containing the complete project context is built.

Within prepare_or_run_pipeline() a yaml file is created as an intermediate representation of the pipeline and uploaded to the Kubeflow instance. To create this yaml file a callable is defined within which a dsl.ContainerOp is created for each step. This ContainerOp contains the container entrypoint command and arguments that will make the image run just the one step. The ContainerOps are assembled according to their interdependencies inside a dsl.Pipeline which can then be compiled into the yaml file.

Handling per-step resources

If your orchestrator allows specification of per-step resources, make sure to handle the configurations defined on each step:

def prepare_or_run_pipeline(...):
    for step in deployment.steps.values():
        if self.requires_resources_in_orchestration_environment(step):
            # make sure to handle the specified resources
            ...

Base Implementation of the Step Entrypoint Configuration

Within the base Docker images that are used for container-based orchestration the src.zenml.entrypoints.entrypoint.py is the default entrypoint to run a specific step. It does so by loading an orchestrator specific StepEntrypointConfiguration object. This object is then used to parse all entrypoint arguments (e.g. --step_name <STEP_NAME>). Finally, the StepEntrypointConfiguration.run() method is used to execute the step. Under the hood this will eventually also call the orchestrators run_step() method.

The StepEntrypointConfiguration is the base class that already implements most of the required functionality. Let’s dive right into it:

1. It defines some mandatory arguments for the step entrypoint. These are set as constants at the top of the file and used as the minimum required arguments.
2. The run() method uses the parsed arguments to set up all required prerequisites before ultimately executing the step.

Here is a schematic view of what the StepEntrypointConfiguration looks like:

from typing import Optional, Set, Any, List

from zenml.entrypoints.base_entrypoint_configuration import (
    BaseEntrypointConfiguration,
)

STEP_NAME_OPTION = "step_name"

class StepEntrypointConfiguration(BaseEntrypointConfiguration):
    """Base class for entrypoint configurations that run a single step."""

    def post_run(
        self,
        pipeline_name: str,
        step_name: str,
    ) -> None:
        """Does cleanup or post-processing after the step finished running."""

    @classmethod
    def get_entrypoint_options(cls) -> Set[str]:
        """Gets all options required for running with this configuration."""
        return super().get_entrypoint_options() | {STEP_NAME_OPTION}

    @classmethod
    def get_entrypoint_arguments(
        cls,
        **kwargs: Any,
    ) -> List[str]:
        """Gets all arguments that the entrypoint command should be called with."""
        return super().get_entrypoint_arguments(**kwargs) + [
            f"--{STEP_NAME_OPTION}",
            kwargs[STEP_NAME_OPTION],
        ]

    def run(self) -> None:
        """Prepares the environment and runs the configured step."""
        ...

This is a slimmed-down version of the base implementation which aims to highlight the abstraction layer. In order to see the full implementation and get the complete docstrings, please check the source code on GitHub.

Build your own Step Entrypoint Configuration

If you need to customize what happens when a step gets executed inside the entrypoint, you can subclass from the StepEntrypointConfiguration class:

If you want to provide a custom run name (this has to be the same for all steps that are executed as part of the same pipeline run), you can overwrite the get_run_name(...) method.

If you need to pass additional arguments to the entrypoint, there are two methods that you need to implement:

• get_entrypoint_options(): This method should return all the additional options that you require in the entrypoint.
• get_entrypoint_arguments(...): This method should return a list of arguments that should be passed to the entrypoint. The arguments need to provide values for all options defined in the get_entrypoint_options() method mentioned above.

Enabling CUDA for GPU-backed hardware

Note that if you wish to use your custom 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.