Runtime Settings

Configuring pipelines, steps, and stack components in ZenML.

Video Tutorial

Configuring pipelines, steps, and stack components in ZenML
This video gives an overview of everything discussed in this chapter, especially with a focus on the post ZenML 0.20.0 world!

Settings in ZenML

As discussed in a previous chapter, there are two ways to configure anything in ZenML:
  • BaseParameters: Runtime configuration passed down as a parameter to step functions.
  • BaseSettings: Runtime settings passed down to stack components and pipelines.
We have already discussed BaseParameters and now is the time to talk about its brother, BaseSettings.

What can be configured?

Looked at one way, BaseParameters configure steps within a pipeline to behave in a different way during runtime. But what other things can be configured at runtime? Here is a list:
  • The resources of 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.
Stack Component Config vs Settings in ZenML

Using objects or dicts

Settings can be passed in directly as BaseSettings-subclassed objects, or a dict-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']}}

How to use 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:
Once you set settings on a pipeline, they will be applied to all steps with some exception. 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:
def my_step() -> None:
print("my step")
def my_pipeline(step1):
# Same as passing it in the step decorator
step_instance = my_step().configure(settings=...)
pipeline_instance = my_pipeline(
step1 = step_instance
# Same as passing it in the pipeline decorator
# Or you can pass it in the run function

Method 3: Configuring with YAML

As all settings can be passed through as a dict, users have the option to send all configuration 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 in the run method of a pipeline instance:
def my_step() -> None:
print("my step")
def my_pipeline(step1):
pipeline_instance = my_pipeline(
step1 = my_step()
# Pass in a config file
The format of a YAML config file is exactly the same as the dict you would pass in python in the above two sections. The step specific settings are nested in a key called steps. Here is rough skeleton of a valid YAML config. All keys are optional.
enable_cache: True
tags: production
run_name: my_run
schedule: {}
settings: {} # same as pipeline settings
settings: {} # same as step settings
settings: {}
ZenML provides a convenient method that takes a pipeline instance and generates a config template based on its settings automatically:
This will write a template file at /local/path/to/config.yaml with a commented out YAML file with all possible options that the pipeline instance can take.
Here is an example of a YAML config file generated from the above method:
An example of a YAML config

The extra dict

You might have noticed another dict that is available to pass through to steps and pipelines called extra. This dict 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 with the post-execution workflow:
from zenml.post_execution import get_pipeline
p = get_pipeline('my_pipeline')
# print out the 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 class/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:
step_instance = my_step()
step_instance.configure(settings={"resources": ResourceSettings(gpu_count=1, memory="2GB")})
step_instance.configuration.settings["resources"] # cpu_count: 2, gpu_count=1, memory="2GB"
In the above example, the two settings were automatically merged.