There are many reasons that you should be using kedro. If you are on a team of Data Scientists/Data Engineers processing DataFrames from many data sources should be considering a pipeline framework. Kedro is a great option that provides many benefits for teams to collaborate, develop, and deploy data pipelines

what-is-kedro

Starter Template

Kedro makes it super easy to get started with their cli that utilizes cookiecutter under the hood.


conda create -n my-new-project -y python=3.8
kedro new
kedro install
kedro run

create-new-kedro-project

read more about how to start your first kedro project here

Collaboration

Kedro provides many tools that help teams collaborate on a single codebase. While writing monolithic scripts it can be easy to pin yourself in a corner where it is difficult to have multiple people making changes to the notebook/script at the same time. Kedro helps guide your team to break your project down into small pieces that different members of the team can work on in parallel.

sharable catalog

Kedro makes it easy to collaborate with members who aren't even working on the pipeline. I often see team members who want to investigate datasets from different points in the pipeline. Kedro makes it really easy for them to load it into python.

for python users

Share catalog entries with folks doing EDA.


catalog.load('main_table')

for non-python users

For those who may not be using python, we can easily kick out a CSV version of that main_table that they can get from s3 or your cloud storage solution of choice.


master_table:
  type: pandas.CSVDataSet
  filepath: s3://bucket/data/03_primary/master_table.csv
  layer: primary

for the SQL folks

We aren't even constrained to those who only use python or excel, we can kick out any kind of dataset that python can output. Kedro even comes with many DataSet types out of the box so that we don't have to write any read/write code.


master_table:
  type: SQLTableDataSet
  table_name: master_table
  credentials: postgres

small nodes over monolithic scripts

As I said before single notebooks/scripts are really hard to collaborate on. I have seen Data Engineers sitting idle waiting to get their changes manually added into the master notebook. When you find yourself in this situation, find a better solution. It's time to break things down into individual modules and utilize a version control system that can automatically merge changes in.

Kedro encourages the use of git version control and storing all node functions inside of modules while still making it really easy to load data into a notebook/shell and start trying out new things.

No More read and write code

As I said earlier kedro comes with datasets for the most popular output formats. It is also backed by a really amazing library called fsspec, this library makes the filesystem that you are storing agnostic to how you write to it. This means that the kedro library utilizes fsspec under the hood and writes to the file as if it was to disk, but based on the prefix to the file it may actually be writing to the local filesystem, gcp, azure blob, or s3.

custom DataSets

If kedro does not have a DataSet for the format that you need to read or write you can easily create your own custom DataSet all you need to do is inherit from kedro.io.AbstractDataSet and create methods for __init__, _load, _save, _exists, and _describe.

Check out this example from their docs. I removed the docstrings for brevity, you can see the entire DataSet in their docs.

The complete example all in one was only available in an older version, more up to date docs have a good writeup that walks through everything separately.


from os.path import isfile
from typing import Any, Union, Dict

import pandas as pd

from kedro.io import AbstractDataSet

class ExcelLocalDataSet(AbstractDataSet):

    def _describe(self) -> Dict[str, Any]:
        return dict(filepath=self._filepath,
                    engine=self._engine,
                    load_args=self._load_args,
                    save_args=self._save_args)

    def __init__(
        self,
        filepath: str,
        engine: str = "xlsxwriter",
        load_args: Dict[str, Any] = None,
        save_args: Dict[str, Any] = None,
    ) -> None:

        self._filepath = filepath
        default_save_args = {}
        default_load_args = {"engine": "xlrd"}

        self._load_args = {**default_load_args, **load_args} \
            if load_args is not None else default_load_args
        self._save_args = {**default_save_args, **save_args} \
            if save_args is not None else default_save_args
        self._engine = engine

    def _load(self) -> Union[pd.DataFrame, Dict[str, pd.DataFrame]]:
        return pd.read_excel(self._filepath, **self._load_args)

    def _save(self, data: pd.DataFrame) -> None:
        writer = pd.ExcelWriter(self._filepath, engine=self._engine)
        data.to_excel(writer, **self._save_args)
        writer.save()

    def _exists(self) -> bool:
        return isfile(self._filepath)

Execution order is taken care of

As you build up complex pipelines containing 10's or 100's of nodes it becomes difficult to splice in new nodes/steps without messing up or a framework to help. Kedro simply needs a set of nodes that each takes in catalog entries as input and output to catalog entries and it will figure out the order for you.

These nodes can be made for one-off purposes, take in functions from reusable libraries, and even be dynamically generated from a configuration. There is no need to worry about hand curating the execution order, that's all taken care of.

Easily slice up a pipeline

Since kedro is a DAG that takes in a pile of nodes and figures out all of the dependencies for you it knows a lot about your pipeline. You can slice it up to only the specific pieces that you need.


# single nodes
pipeline.only_nodes("node1")

# single nodes and all of thier dependencies
pipeline.to_nodes("node1", "node2")

# from a dataset to all of its dependants
pipeline.from_inputs("dataset1", "dataset2")

# to a an outputs with all of its dependencies
pipeline.to_outputs("dataset6", "dataset7")

plugins/hooks

Creating your own modifications to how kedro behaves is made really simple through the use of hooks. There are several hooks that happen at different points in the kedro lifecycle. For instance, you can hook in before pipeline run or after pipeline run to do whatever your project needs.

creating-the-kedro-preflight-hook

pip install plugin

There is a growing list of plugins available from pypi that is only a pip install away. Most of them are on GitHub and tagged as a kedro-plugin topic.

flexible cli

In the end, you have a cli for your project that can run your pipeline in all sorts of cool ways since it knows about each node's dependencies. This makes running and scheduling production a breeze.


# single nodes
kedro run --node node1

# single nodes and all of their dependencies
kedro run --to-nodes node1,node2

# from a dataset to all of its dependents
kedro run --from-inputs dataset1,dataset2

# to outputs with all of their dependencies
kedro run --to-outputs dataset6,dataset7

Try it out

Hopefully this post gave you the inspiration to get started today, if it did pip install kedro and run kedro new to try it out.