Frequently asked questions¶
Note: This documentation is based on
Kedro 0.15.5, if you spot anything that is incorrect then please create an issue or pull request.
What is Kedro?¶
Kedro is a workflow development tool that helps you build data pipelines that are robust, scaleable, deployable, reproducible and versioned. It was originally designed by Aris Valtazanos and Nikolaos Tsaousis at QuantumBlack to solve the challenges they faced in their project work.
This work was later turned into a product thanks to the following contributors: Ivan Danov, Dmitrii Deriabin, Gordon Wrigley, Yetunde Dada, Nasef Khan, Kiyohito Kunii, Nikolaos Kaltsas, Meisam Emamjome, Peteris Erins, Lorena Balan, Richard Westenra and Anton Kirilenko.
What are the primary advantages of Kedro?¶
It is important to consider the primary advantages of Kedro over existing tools.
As we see it, Kedro emphasises a seamless transition from development to production without slowing the pace of the experimentation stage, because it:
- Simplifies data access, using YAML configuration to define a single-source of truth for all data sources that your workflow requires
- Uses a familiar data interface, by borrowing arguments from Pandas and Spark APIs meaning you do not have to learn a new API
- Has a minimal pipeline syntax, that uses Python functions
- Makes datasets 1st-level citizens, resolving task running order according to what each task produces and consumes, meaning you do not need to explicitly define dependencies between tasks
- Has built-in runner selection, choosing sequential or parallel runner functionality is a
- Has a low-effort setup, that does not need a scheduler or database
- Starts with a project template, which has built-in conventions and best practices from 50+ analytics engagements
- Is flexible, simplifying your extension or replacement of core functionality e.g. the whole Data Catalog could be replaced with another mechanism for data access like
How does Kedro compare to other projects?¶
Data pipelines consist of extract-transform-load (ETL) workflows. If we understand that data pipelines must be scaleable, monitored, versioned, testable and modular then this introduces us to a spectrum of tools that can be used to construct such data pipelines.
Pipeline abstraction is implemented in workflow schedulers like Luigi and Airflow, as well as in ETL frameworks like Bonobo ETL and Bubbles.
Kedro vs workflow schedulers¶
Kedro is not a workflow scheduler like Airflow and Luigi. Kedro makes it easy to prototype your data pipeline, while Airflow and Luigi are complementary frameworks that are great at managing deployment, scheduling, monitoring and alerting. A Kedro pipeline is like a machine that builds a car part. Airflow and Luigi tell the different Kedro machines to switch on or off in order to work together to produce a car. We have built a Kedro-Airflow plugin, providing faster prototyping time and reducing the barriers to entry associated with moving pipelines to Airflow.
Kedro vs other ETL frameworks¶
The primary differences to Bonobo ETL and Bubbles are related to the following features of Kedro:
- Ability to support big data operations. Kedro supports big data operations by allowing you to use PySpark on your projects. We also look at processing dataframes differently to both tools as we consider entire dataframes and do not make use of the slower line-by-line data stream processing.
- Project structure. Kedro provides a built-in project structure from the beginning of your project configured for best-practice project management.
- Automatic dependency resolution for pipelines. The
Pipelinemodule also maps out dependencies between nodes and displays the results of this in a sophisticated but easy to understand directed acyclic graph.
Where can I learn more?¶
You have two sources of community-generated content:
- Articles, podcasts and talks
- Kedro used on real-world use cases
We’ll be updating links to community-generated content when we come across them.
Articles, podcasts and talks¶
|Using Kedro and MLflow Deploying and versioning data pipelines at scale||Tom Goldenberg & Musa Bilal||QuantumBlack Medium||Article||https://medium.com/@QuantumBlack/deploying-and-versioning-data-pipelines-at-scale-942b1d81b5f5|
|Kedro: A New Tool For Data Science||Jo Stichbury||Towards Data Science||Article||https://towardsdatascience.com/kedro-prepare-to-pimp-your-pipeline-f8f68c263466|
|Building a Pipeline with Kedro for an ML Competition||Masaaki Hirotsu||mhiro2||Article||https://medium.com/mhiro2/building-pipeline-with-kedro-for-ml-competition-63e1db42d179|
|Kedro in Jupyter Notebooks On Google GCP Dataproc||Zhong Chen||Zhong Chen||Article||https://medium.com/@zhongchen/kedro-in-jupyter-notebooks-on-google-gcp-dataproc-31d5f45ad235|
|Ship Faster With An Opinionated Data Pipeline Framework Episode 100||Tom Goldenberg||Data Engineering Podcast||Podcast||https://www.dataengineeringpodcast.com/kedro-data-pipeline-episode-100/|
|Production-level data pipelines that make everyone happy using Kedro||Yetunde Dada||PyData Berlin 2019||Recording||https://youtu.be/OFObles2CJs|
|Data Science Best Practices con Kedro||Carlos Gimenez||PyData Córdoba Argentina 2019||Recording||https://youtu.be/_0kMmRfltEQ|
|Kedro - Nubank ML Meetup||Carlos Barreto||Nubank On the Stage||Recording||https://youtu.be/clBgxmDsSjI|
What is data engineering convention?¶
Note: The data layers don’t have to exist locally in the
datafolder within your project. It is recommended that you structure your S3 buckets or other data stores in a similar way.
|Folder in data||Description|
|Raw||Initial start of the pipeline, containing the sourced data model(s) that should never be changed, it forms your single source of truth to work from. These data models are typically un-typed in most cases e.g. csv, but this will vary from case to case.|
|Intermediate||Optional data model(s), which are introduced to type your
|Primary||Domain specific data model(s) containing cleansed, transformed and wrangled data from either
|Feature||Analytics specific data model(s) containing a set of features defined against the
|Model input||Analytics specific data model(s) containing all
|Models||Stored, serialised pre-trained machine learning models.|
|Model output||Analytics specific data model(s) containing the results generated by the model based on the
|Reporting||Reporting data model(s) that are used to combine a set of
What version of Python does Kedro use?¶
Kedro is built for Python 3.5, 3.6 and 3.7.
How do I upgrade Kedro?¶
We use Semantic Versioning. The best way to safely upgrade is to check our release notes for any notable breaking changes. Please also follow the steps in the migration guide if one is included for a particular release.
Once Kedro is installed, you can check your version as follows:
To later upgrade Kedro to a different version, simply run:
pip install kedro -U
When migrating an existing project to a newer Kedro version, make sure you also update the
project_version in your
ProjectContext, which is found in
What best practice should I follow to avoid leaking confidential data?¶
- Avoid committing data to version control (data folder is by default ignored via
- Avoid committing data to notebook output cells (data can easily sneak into notebooks when you don’t delete output cells)
- Don’t commit sensitive results or plots to version control (in notebooks or otherwise)
- Don’t commit credentials in
conf/. There are two default folders for adding configuration -
conf/local/. Only the
conf/local/folder should be used for sensitive information like access credentials. To add credentials, please refer to the
conf/base/credentials.ymlfile in the project template.
- By default any file inside the
conf/folder (and its subfolders) containing
credentialsin its name will be ignored via
.gitignoreand not committed to your git repository.
- To describe where your colleagues can access the credentials, you may edit the
README.mdto provide instructions.
What is the philosophy behind Kedro?¶
Kedro is a Python library and lightly opinionated framework. This means that we give you the flexibility and extensibility of a standard Python library and make very few assumptions on the best way to do things. We have created independent but friendly modules – modules that understand each others’ defaults and are compatible. You can use alternative methods and choose to use one or all of the modules but it is understood that using Kedro in its entirety is the best thing that you can do for your projects.
The Kedro design principles are:
- Declarative definitions
Where do I store my custom editor configuration?¶
You can use
conf/local to describe your custom editor configuration.
How do I look up an API function?¶
Every Kedro function or class has extensive help, so please do take advantage of this capability, example of this is presented below:
How do I build documentation for my project?¶
Project-specific documentation can be generated by running
kedro build-docs in the project root directory. This will create documentation based on the code structure. Documentation will also include the
docstrings defined in the project code.
HTML files for the project documentation will be built to
How do I build documentation about Kedro?¶
A local copy of documentation about Kedro can be generated by running
kedro docs from the command line. The documentation is also available online.