Kedro is an unopinionated Data Engineering framework that comes with a somewhat
opinionated template. It gives the user a way to build pipelines that
automatically take care of io through the use of abstract
DataSets that the
user specifies through
Catalog entries. These
Catalog entries are loaded,
ran through a function, and saved by
Nodes. The order that these
executed are determined by the
Pipeline, which is a DAG. It's the
runner's job to manage the execution of the
This is an updated version of my original what-is-kedro article
If you are doing a series of operations to data with python, especially if you are using something as supported as pandas, you should be using a framework that gives you a pipeline as a DAG and abstracts io.
Like I said,
kedro is unopinionated it does determine where or how your data
should be ran. The kedro team does support the following Orchestrators
with very little add on to the base template.
Did I say kedro is unopionated? Datasets are what allow kedro too be so
flexible accross a number of different python objects. Any python object can
be made into a kedro dataset. Kedro comes out of the box with many purpose built
DataSets like storing pandas DataFrames to parquet, csv, or a sql table. If
kedro does not come with support for the type of python objects you work with
don't worry, you can for the closest option they support and build your own.
Or if you do not want to build your own, you can use a
You will not often be creating your own datasets, most of what you need whould
already be taken care of by the kedro framework. What you will need to do is
to use the existing
DataSets to build your data catalog.
Kedro takes care of all fo the file io for you, you simply need to use the
catalog to tell kedro what type of DataSet to use and any extra information
DataSet needs. Much of the time this is simply a filepath.
Typically the catalog is specified in yaml format. If you are not familiar with yaml, I suggest learnxinyminutes.com/docs/yaml/ as a resource of examples.
test: type: pandas.CSVDataSet filepath: s3://your_bucket/test.csv #
Here is the most basic yaml catalog entry taken from the kedro docs
cars: type: pandas.CSVDataSet filepath: data/01_raw/company/cars.csv sep: ',' load_args: save_args: index: False date_format: '%Y-%m-%d %H:%M' decimal: .
Here is a bit more complex example that takes in
Nodes are a very core part of kedro to build the DAG. These nodes are what provides the definition of what catalog entries, get passed into which function, and output to another catalog entry.
import pandas as pd import numpy as np def clean_data(cars: pd.DataFrame, boats: pd.DataFrame) -> Dict[str, pd.DataFrame]: return dict(cars_df=cars.dropna(), boats_df=boats.dropna()) def halve_dataframe(data: pd.DataFrame) -> List[pd.DataFrame]: return np.array_split(data, 2) nodes = [ node(clean_data, inputs=['cars2017', 'boats2017'], outputs=dict(cars_df='clean_cars2017', boats_df='clean_boats2017')), node(halve_dataframe, 'clean_cars2017', ['train_cars2017', 'test_cars2017']), node(halve_dataframe, dict(data='clean_boats2017'), ['train_boats2017', 'test_boats2017']) ]
Here is an example of three nodes taken from their docs
Pipeline, is a DAG (Directed Acyclic Graph). It is a graph object
that flows in one direction. You can slice into the pipeline using a few built
in graph method
can chain up these method calls since each one returns a new
You can also ask a pipline for its edges with
outputs. You can
also list every dataset along the way with
Lastly you can convert it back into a list of nodes with
from kedro.pipeline import Pipeline, node # using our nodes from last tim Pipeline(nodes)
The runner is the bridge between kedro and the orchestrators. The kedro team provides some basic runners for running pipelines locally, built right into the framework, but adding on new runners for different orchestrators is done through the use of adding in a new runner to your project.
Kedro allows you to hook into a number of lifecycle methods through the use of
pluggy framework. Yes the one that
pytest is built on. There are a
number of different lifecycle methods that allow us to hook in around where
kedro is running such as