0.46.0
Cloud
Product
Resources
GitHub

Configure steps/pipelines

Configuring pipelines, steps, and stack components in ZenML.

This is an older version of the ZenML documentation. To read and view the latest version please visit this up-to-date URL.

Configure steps/pipelines

Define steps

To define a step, all you need to do is decorate your Python functions with ZenML's @step decorator:

from zenml import step

@step
def my_function():
    ...

Type annotations

Your functions will work as ZenML steps even if you don't provide any type annotations for their inputs and outputs. However, adding type annotations to your step functions gives you lots of additional benefits:

  • Type validation of your step inputs: ZenML makes sure that your step functions receive an object of the correct type from the upstream steps in your pipeline.

  • Better serialization: Without type annotations, ZenML uses Cloudpickle to serialize your step outputs. When provided with type annotations, ZenML can choose a materializer that is best suited for the output. In case none of the builtin materializers work, you can even write a custom materializer.

from typing import Tuple
from zenml import step

@step
def square_root(number: int) -> float:
    return number ** 0.5

# To define a step with multiple outputs, use a `Tuple` type annotation
@step
def divide(a: int, b: int) -> Tuple[int, int]:
    return a // b, a % b

It is impossible for ZenML to detect whether you want your step to have a single output artifact of type Tuple or multiple output artifacts just by looking at the type annotation.

We use the following convention to differentiate between the two: When the return statement is followed by a tuple literal (e.g. return 1, 2 or return (value_1, value_2)) we treat it as a step with multiple outputs. All other cases are treated as a step with a single output of type Tuple.

# Single output artifact
@step
def my_step() -> Tuple[int, int]:
    output_value = (0, 1)
    return output_value

# Single output artifact with variable length
@step
def my_step(condition) -> Tuple[int, ...]:
    if condition:
        output_value = (0, 1)
    else:
        output_value = (0, 1, 2)

    return output_value

# Single output artifact using the `Annotated` annotation
@step
def my_step() -> Annotated[Tuple[int, ...], "my_output"]:
    return 0, 1


# Multiple output artifacts
@step
def my_step() -> Tuple[int, int]:
    return 0, 1


# Not allowed: Variable length tuple annotation when using
# multiple output artifacts
@step
def my_step() -> Tuple[int, ...]:
    return 0, 1

If you want to make sure you get all the benefits of type annotating your steps, you can set the environment variable ZENML_ENFORCE_TYPE_ANNOTATIONS to True. ZenML will then raise an exception in case one of the steps you're trying to run is missing a type annotation.

Step output names

By default, ZenML uses the output name output for single output steps and output_0, output_1, ... for steps with multiple outputs. These output names are used to display your outputs in the dashboard and fetch them after your pipeline is finished.

If you want to use custom output names for your steps, use the Annotated type annotation:

from typing_extensions import Annotated  # or `from typing import Annotated on Python 3.9+
from typing import Tuple
from zenml import step

@step
def square_root(number: int) -> Annotated[float, "custom_output_name"]:
    return number ** 0.5

@step
def divide(a: int, b: int) -> Tuple[
    Annotated[int, "quotient"],
    Annotated[int, "remainder"]
]:
    return a // b, a % b

Configure steps/pipelines

Parameters for your steps

When calling a step in a pipeline, the inputs provided to the step function can either be an artifact or a parameter. An artifact represents the output of another step that was executed as part of the same pipeline and serves as a means to share data between steps. Parameters, on the other hand, are values provided explicitly when invoking a step. They are not dependent on the output of other steps and allow you to parameterize the behavior of your steps.

In order to allow the configuration of your steps using a configuration file, only values that can be serialized to JSON using Pydantic can be passed as parameters.

@step
def my_step(input_1: int, input_2: int) -> None:
    pass


@pipeline
def my_pipeline():
    int_artifact = some_other_step()
    # We supply the value of `input_1` as an artifact and
    # `input_2` as a parameter
    my_step(input_1=int_artifact, input_2=17)
    # We could also call the step with two artifacts or two
    # parameters instead:
    # my_step(input_1=int_artifact, input_2=int_artifact)
    # my_step(input_1=1, input_2=2)

Parameters and caching

When an input is passed as a parameter, the step will only be cached if all parameter values are exactly the same as for previous executions of the step.

Artifacts and caching

When an artifact is used as a step function input, the step will only be cached if all the artifacts are exactly the same as for previous executions of the step. This means that if any of the upstream steps that produce the input artifacts for a step were not cached, the step itself will always be executed.

Pass any kind of data to your steps

External artifacts can be used to pass values to steps that are neither JSON serializable nor produced by an upstream step.

import numpy as np
from zenml import pipeline, step
from zenml.artifacts.external_artifact import ExternalArtifact

@step
def trainer(data: np.ndarray) -> ...:
    ...

@pipeline
def my_pipeline():
    trainer(data=ExternalArtifact(np.array([1, 2, 3])))

    # This would not work, as only JSON serializable values
    # or artifacts can be passed as a step input:
    # trainer(data=np.array([1, 2, 3]))

Optionally, you can configure the ExternalArtifact to use a custom materializer for your data or disable artifact metadata and visualizations. Check out the SDK docs for all available options.

Using an ExternalArtifact with input data for your step automatically disables caching for the step.

You can also use an ExternalArtifact to pass an artifact stored in the ZenML database. Search can be performed using the UUID of an artifact:

from uuid import UUID

artifact = ExternalArtifact(id=UUID("3a92ae32-a764-4420-98ba-07da8f742b76"))

Another way to search would be by a combination of a pipeline name where the artifact was generated and the artifact name itself. If you search by pipeline/artifact name pair the search will happen using the last successful run of the pipeline:

artifact = ExternalArtifact(pipeline_name="training_pipeline", artifact_name="model")
See it in action with the E2E example

To setup the local environment used below, follow the recommendations from the Project templates.

In pipelines/batch_inference.py, you can find an example using the ExternalArtifact concept to share Artifacts produced by a training pipeline inside a batch inference pipeline.

On the ETL stage pipeline, developers can pass a sklearn.Pipeline fitted during training for feature preprocessing and apply it to transform inference input features. With this, we ensure that the exact same feature preprocessor used during training will be used during inference.

    ########## ETL stage  ##########
    df_inference, target = data_loader(is_inference=True)
    df_inference = inference_data_preprocessor(
        dataset_inf=df_inference,
        preprocess_pipeline=ExternalArtifact(
            pipeline_name=MetaConfig.pipeline_name_training,
            artifact_name="preprocess_pipeline",
        ),
        target=target,
    )

On the DataQuality stage pipeline, developers can pass pd.DataFrame used as a training dataset to be used as a reference dataset versus the current inference one to apply Evidently and get DataQuality report back. With this, we ensure that the exact same training dataset used during the training phase will be used to compare with the inference dataset here.

    ########## DataQuality stage  ##########
    report, _ = evidently_report_step(
        reference_dataset=ExternalArtifact(
            pipeline_name=MetaConfig.pipeline_name_training,
            artifact_name="dataset_trn",
        ),
        comparison_dataset=df_inference,
        ignored_cols=["target"],
        metrics=[
            EvidentlyMetricConfig.metric("DataQualityPreset"),
        ],
    )

Using a custom step invocation ID

When calling a ZenML step as part of your pipeline, it gets assigned a unique invocation ID that you can use to reference this step invocation when defining the execution order of your pipeline steps or use it to fetch information about the invocation after the pipeline has finished running.

from zenml import pipeline, step

@step
def my_step() -> None:
    ...

@pipeline
def example_pipeline():
    # When calling a step for the first time inside a pipeline,
    # the invocation ID will be equal to the step name -> `my_step`.
    my_step()
    # When calling the same step again, the suffix `_2`, `_3`, ... will
    # be appended to the step name to generate a unique invocation ID.
    # For this call, the invocation ID would be `my_step_2`.
    my_step()
    # If you want to use a custom invocation ID when calling a step, you can
    # do so by passing it like this. If you pass a custom ID, it needs to be
    # unique for all the step invocations that happen as part of this pipeline.
    my_step(id="my_custom_invocation_id")

Control the execution order

By default, ZenML uses the data flowing between steps of your pipeline to determine the order in which steps get executed.

The following example shows a pipeline in which step_3 depends on the outputs of step_1 and step_2. This means that ZenML can execute both step_1 and step_2 in parallel but needs to wait until both are finished before step_3 can be started.

from zenml import pipeline

@pipeline
def example_pipeline():
    step_1_output = step_1()
    step_2_output = step_2()
    step_3(step_1_output, step_2_output)

If you have additional constraints on the order in which steps get executed, you can specify non-data dependencies by passing the invocation IDs of steps that should run before your step like this: my_step(after="other_step"). If you want to define multiple upstream steps, you can also pass a list for the after argument when calling your step: my_step(after=["other_step", "other_step_2"]).

Check out the previous section to learn about the invocation ID and how to use a custom one for your steps.

from zenml import pipeline

@pipeline
def example_pipeline():
    step_1_output = step_1(after="step_2")
    step_2_output = step_2()
    step_3(step_1_output, step_2_output)

This pipeline is similar to the one explained above, but this time ZenML will make sure to only start step_1 after step_2 has finished.

Enable or disable logs storing

By default, ZenML uses a special logging handler to capture the logs that occur during the execution of a step. These logs are stored within the respective artifact store of your stack.

import logging

from zenml import step

@step 
def my_step() -> None:
    logging.warning("`Hello`")  # You can use the regular `logging` module.
    print("World.")  # You can utilize `print` statements as well.

You can display the logs in the dashboard as follows:

If you do not want to store the logs in your artifact store, you can:

  1. Disable it by using the enable_step_logs parameter either with your @pipeline or @step decorator:

    from zenml import pipeline, step

@step(enable_step_logs=False)  # disables logging for this step
def my_step() -> None:
    ...

@pipeline(enable_step_logs=False)  # disables logging for the entire pipeline
def my_pipeline():
    ...

  2. Disable it by using the environmental variable ZENML_DISABLE_STEP_LOGS_STORAGE. This environmental variable takes precedence over the parameters mentioned above.

Settings in ZenML

Settings in ZenML allow you to configure runtime configurations for stack components and pipelines. Concretely, they allow you to configure:

  • The resources required for a step

  • Configuring the containerization process of a pipeline (e.g. What requirements get installed in the Docker image)

  • Stack component-specific configuration, e.g., if you have an experiment tracker passing in the name of the experiment at runtime

You will learn about all of the above in more detail later, but for now, let's try to understand that all of this configuration flows through one central concept, called BaseSettings (From here on, we use settings and BaseSettings as analogous in this guide).

Types of settings

Settings are categorized into two types:

For stack-component-specific settings, you might be wondering what the difference is between these and the configuration passed in while doing zenml stack-component register <NAME> --config1=configvalue --config2=configvalue, etc. The answer is that the configuration passed in at registration time is static and fixed throughout all pipeline runs, while the settings can change.

A good example of this is the MLflow Experiment Tracker, where configuration which remains static such as the tracking_url is sent through at registration time, while runtime configuration such as the experiment_name (which might change every pipeline run) is sent through as runtime settings.

Even though settings can be overridden at runtime, you can also specify default values for settings while configuring a stack component. For example, you could set a default value for the nested setting of your MLflow experiment tracker: zenml experiment-tracker register <NAME> --flavor=mlflow --nested=True

This means that all pipelines that run using this experiment tracker use nested MLflow runs unless overridden by specifying settings for the pipeline at runtime.

Using objects or dicts

Settings can be passed in directly as BaseSettings-subclassed objects, or a dictionary representation of the object. For example, a Docker configuration can be passed in as follows:

from zenml.config import DockerSettings

settings = {'docker': DockerSettings(requirements=['pandas'])}

Or like this:

settings = {'docker': {'requirements': ['pandas']}}

Utilizing the settings

Method 1: Directly on the decorator

The most basic way to set settings is through the settings variable that exists in both @step and @pipeline decorators:

@step(settings=...)


...


@pipeline(settings=...)


...

Once you set settings on a pipeline, they will be applied to all steps with some exceptions. See the later section on precedence for more details.

Method 2: On the step/pipeline instance

This is exactly the same as passing it through the decorator, but if you prefer you can also pass it in the configure methods of the pipeline and step instances:

@step
def my_step() -> None:
    print("my step")


@pipeline
def my_pipeline():
    my_step()


# Same as passing it in the step decorator
my_step.configure(settings=...)

# Same as passing it in the pipeline decorator
my_pipeline.configure(settings=...)

Method 3: Configuring with YAML

As all settings can be passed through as a dictionary, users have the option to send all configurations in via a YAML file. This is useful in situations where code changes are not desirable.

To use a YAML file, you must pass it to the with_options(...) method of a pipeline:

@step
def my_step() -> None:
    print("my step")


@pipeline
def my_pipeline():
    my_step()


# Pass in a config file
my_pipeline = my_pipeline.with_options(config_path='/local/path/to/config.yaml')

The format of a YAML config file is exactly the same as the configurations you would pass in Python in the above two sections. Step-specific configurations can be passed by using the step invocation ID inside the steps dictionary. Here is a rough skeleton of a valid YAML config. All keys are optional.

enable_cache: True
enable_artifact_metadata: True
enable_artifact_visualization: True
extra:
  tags: production
run_name: my_run
schedule: { }
settings: { }  # same as pipeline settings
steps:
  step_invocation_id:
    settings: { }  # same as step settings
  other_step_invocation_id:
    settings: { }
  ...

You can also use the following method to generate a config template (at path /local/path/to/config.yaml) that includes all configuration options for this specific pipeline and your active stack:

my_pipeline.write_run_configuration_template(path='/local/path/to/config.yaml')

Here is an example of a YAML config file generated from the above method:

An example of a YAML config

Some configuration is commented out as it is not needed.

enable_cache: True
enable_artifact_metadata: True
enable_artifact_visualization: True
extra:
  tags: production
run_name: my_run
# schedule:
#   catchup: bool
#   cron_expression: Optional[str]
#   end_time: Optional[datetime]
#   interval_second: Optional[timedelta]
#   start_time: Optional[datetime]
settings:
  docker:
    build_context_root: .
    # build_options: Mapping[str, Any]
    # source_files: str
    # copy_global_config: bool
    # dockerfile: Optional[str]
    # dockerignore: Optional[str]
    # environment: Mapping[str, Any]
    # install_stack_requirements: bool
    # parent_image: Optional[str]
    # replicate_local_python_environment: Optional
    # required_integrations: List[str]
    requirements:
      - pandas
    # target_repository: str
    # user: Optional[str]
  resources:
    cpu_count: 1
    gpu_count: 1
    memory: "1GB"
steps:
  get_first_num:
    enable_cache: false
    experiment_tracker: mlflow_tracker
#     extra: Mapping[str, Any]
#     outputs:
#       first_num:
#         artifact_source: Optional[str]
#         materializer_source: Optional[str]
#     parameters: {}
#     settings:
#       resources:
#         cpu_count: Optional[PositiveFloat]
#         gpu_count: Optional[PositiveInt]
#         memory: Optional[ConstrainedStrValue]
#     step_operator: Optional[str]
#   get_random_int:
#     enable_cache: Optional[bool]
#     experiment_tracker: Optional[str]
#     extra: Mapping[str, Any]
#     outputs:
#       random_num:
#         artifact_source: Optional[str]
#         materializer_source: Optional[str]
#     parameters: {}
#     settings:
#       resources:
#         cpu_count: Optional[PositiveFloat]
#         gpu_count: Optional[PositiveInt]
#         memory: Optional[ConstrainedStrValue]
#     step_operator: Optional[str]
#   subtract_numbers:
#     enable_cache: Optional[bool]
#     experiment_tracker: Optional[str]
#     extra: Mapping[str, Any]
#     outputs:
#       result:
#         artifact_source: Optional[str]
#         materializer_source: Optional[str]
#     parameters: {}
#     settings:
#       resources:
#         cpu_count: Optional[PositiveFloat]
#         gpu_count: Optional[PositiveInt]
#         memory: Optional[ConstrainedStrValue]
#     step_operator: Optional[str]

The extra dict

You might have noticed another dictionary that is available to be passed to steps and pipelines called extra. This dictionary is meant to be used to pass any configuration down to the pipeline, step, or stack components that the user has use of.

An example of this is if I want to tag a pipeline, I can do the following:

@pipeline(name='my_pipeline', extra={'tag': 'production'})


...

This tag is now associated and tracked with all pipeline runs, and can be fetched later:

from zenml.client import Client

pipeline_run = Client().get_pipeline_run("<PIPELINE_RUN_NAME>")

# print out the extra
print(pipeline_run.config.extra)
# {'tag': 'production'}

Hierarchy and precedence

Some settings can be configured on pipelines and steps, some only on one of the two. Pipeline-level settings will be automatically applied to all steps, but if the same setting is configured on a step as well that takes precedence. The next section explains in more detail how the step-level settings will be merged with pipeline settings.

Merging settings on instance/run:

When a settings object is configured, ZenML merges the values with previously configured keys. E.g.:

from zenml.config import ResourceSettings


@step(settings={"resources": ResourceSettings(cpu_count=2, memory="1GB")})
def my_step() -> None:
    ...


my_step.configure(
    settings={"resources": ResourceSettings(gpu_count=1, memory="2GB")}
)

my_step.configuration.settings["resources"]
# cpu_count: 2, gpu_count=1, memory="2GB"

In the above example, the two settings were automatically merged.

PreviousPipelining featuresNextCompose pipelines

Last updated