Skip to content

from distributed import Client


Fondant makes it easy to build dataset collaborative leveraging reusable components. Fondant provides a lot of components out of the box (overview), but you can also define your own custom components.

The anatomy of a component#

A component is completely defined by its script, specification, docker image, which data it consumes and produces, and which arguments it takes. The definition of the script is similar for all types of components. All other aspects of the component are defined different ways, depending on the type of component. Continue reading to learn more about the different types of components and how to define them.

Component script#

The logic should be implemented as a class, inheriting from one of the base Component classes offered by Fondant. There are three large types of components:

  • LoadComponent: Load data and initialise a dataset from an external data source
  • TransformComponent: Implement a single transformation step to transform data in your dataset
  • WriteComponent: Write your dataset to an external data sink

The easiest way to implement a TransformComponent is to subclass the provided PandasTransformComponent. This component streams your data and offers it in memory-sized chunks as pandas dataframes.

import pandas as pd
from fondant.component import PandasTransformComponent

class ExampleComponent(PandasTransformComponent):

    def __init__(self, *, argument1, argument2) -> None:
            argumentX: An argument passed to the component
        # Initialize your component here based on the arguments

    def transform(self, dataframe: pd.DataFrame) -> pd.DataFrame:
        """Implement your custom logic in this single method
            dataframe: A Pandas dataframe containing one partition of your data
            A pandas dataframe containing the transformed data

The __init__ method is called once for each component class with custom arguments defined in the args section of the component. This is a good place to initialize resources and costly initializations such as network connections, models, parsing a config file, etc. By doing so, you can effectively prevent the redundant re-initialization of resources each time the transform method is invoked.

The transform method is called multiple times, each time containing a pandas dataframe with a partition of your data loaded in memory.

The dataframes passed to the transform method contains the data specified in the consumes section of the component. If a component defines that it consumes an image field, this data can be accessed using dataframe["image"].

The transform method should return a single dataframe, with the columns complying to the schema defined by the produces section of the component specification.

Configuring Dask#

You can configure the Dask client based on the needs of your component by overriding the dask_client method:

import os

from dask.distributed import Client, LocalCluster
from fondant.component import PandasTransformComponent

class Component(PandasTransformComponent):

    def dask_client(self) -> Client:
        """Initialize the dask client to use for this component."""
        cluster = LocalCluster(
        return Client(cluster)

The default configuration uses a LocalCluster which works with processes, the same amount of workers as logical CPUs available, and one thread per worker.

Some components might work more optimally using threads or a different combination of threads and processes. To use multiple GPUs, you can use a LocalCUDACluster.

Component types#

We can distinguish two different types of components:

  • Custom components are completely defined and implemented by the user. There are two ways to define a custom component:

    • Lightweight Components: Create a component from a self-contained Python function. This is the easiest way to create a custom component. It allows you to define a component without having to build a custom docker image or defining a component specification.
    • Containerized Components: You can build your code into a docker image and write an accompanying component specification that refers to it. This is used for more complex components that require additional dependencies (e.g. GPU support).
  • Reusable components can be used out of the box and can be loaded from the Fondant Hub. They are containerized components that are defined by the Fondant team or the community.

Custom components#

Lightweight Components#

To define a lightweight component, you can create a self-contained python function that implements the logic of your component.

from fondant.component import PandasTransformComponent
from fondant.dataset import  lightweight_component
import pandas as pd
import pyarrow as pa

class AddNumber(PandasTransformComponent):
    def __init__(self, n: int):
        self.n = n

    def transform(self, dataframe: pd.DataFrame) -> pd.DataFrame:
        dataframe["x"] = dataframe["x"].map(lambda x: x + self.n)
        return dataframe

You can apply a custom component to your dataset by passing in the reference to the component class containing your script.
_ = dataset.apply(
    produces={"x": pa.int32()},
      "n": 1

See our best practices on creating a lightweight component.

Containerized Components#

To define your own containerized component, you can build your code into a docker image and write an accompanying component specification that refers to it.

A typical file structure for a custom component looks like this:

|- components
|  |- custom_component
|     |- src
|     |  |-
|     |- Dockerfile
|     |- fondant_component.yaml
|     |- requirements.txt

The Dockerfile is used to build the code into a docker image, which is then referred to in the fondant_component.yaml.

name: Custom component
description: This is a custom component
image: custom_component:latest

You can apply a custom component to your dataset by passing in the path to the directory containing your fondant_component.yaml.
dataset = dataset.apply(
    "arg": "value"

See our best practices on creating a containerized component.

Reusable components#

Reusable components are out of the box containerized components from the Fondant Hub that you can easily add to your dataset:

dataset = dataset.apply(
    "arg": "value"

You can find an overview of the available reusable components on the Fondant hub. Check their documentation for information on which arguments they accept and which data they consume and produce.