Tutorial
5 min read

Running Machine Learning Pipelines with Kedro, Kubeflow and Airflow

One of the biggest challenges of today’s Machine Learning world is the lack of standardization when it comes to models training.

We all know that data needs to be cleaned, split into training and test sets, fitted into the model and validated on the observations from the test subset. Maybe there should be some cross-validation involved, hyperparameters tuning is also not a bad idea. All Data Scientists feel how to work with the models efficiently, but a lack of common standards makes their work hard to understand for other engineers using a different methodology.

To overcome this issue, QuantumBlack open-sourced the Kedro framework, an open-source Python framework for creating reproducible, maintainable and modular data science code. Projects created with Kedro are universal enough to cover most of the tasks that Data Scientists may have. Additionally, Data Catalog and Pipeline abstractions make the model building process look like a software project that can be configured and deployed easily, especially by the engineers that didn’t take part in implementing them. It provides a variety of plugins to log models and metrics in Mlflow, ship the project as a docker image and more. Based on the feedback gathered from our clients, we made Kedro a core part of GetInData Machine Learning Operations Platform. We presented our MLOps platform blueprint during the Google Cloud Region launch in Warsaw. You can watch it here (currently only in Polish)

However, having a Kedro project ready and well tested on the data sample doesn’t mean it is ready enough to go “into production” and be deployed quickly. Continuous training, hyper parameter tuning, continuous quality validation - all these tasks need some scheduling capabilities, distributed computing and powerful hardware to do things efficiently. Kedro describes some ideas on how to deploy the models, but well, you know what they say, “work smarter, not harder - automate everything” ;-)

*Welcome to MLOps candy shop and choose your flavour! - Mateusz Pytel & Mariusz Strzelecki - Big Data Technology Warsaw Summit 2021*

Using Kubeflow? Meet kedro-kubeflow

Kubernetes is the core of our Machine Learning Operations platform and Kubeflow is a system that we often deploy for our clients. Therefore, we decided to automate the generation of the Kubeflow pipeline from the existing Kedro pipeline to allow it to be scheduled by Kubeflow Pipelines (a.k.a. KFP) and started on the Kubernetes cluster. Thankfully, the creators of Kedro gave us a little help, by doing proof-of-concept of this integration and providing interesting insights.

The result of our work is available on GitHub as a kedro-kubeflow plugin. You install it in your existing Kedro project and soon you can:

  • compile Kedro nodes as KFP steps and reflect dependencies,
  • upload the compiled KFP pipeline to the server,
  • trigger the execution using CLI or enable a schedule when you’re finally happy with how the pipeline works,
  • connect seamlessly with MLflow and log all the steps under one Mlflow run (even if they are actually separate processes),
  • access Google AI Platform Pipelines API using IAM Proxy,
  • and more!

We faced some challenges - for example Kedro expects the `data/` directory to be a place where nodes exchange the data, but in a distributed environment there are limited options to maintain shared storage between different processes. Thankfully, with a bit of hacking, we made it! 

getindata-machine-learning-operations-platform-kubeflow-plugin-mlops GetInData MLOps Platform: Kubeflow plugin

Using Airflow? Meet kedro-airflow-k8s

Some of our customers tend to avoid Kubeflow, as the system is quite complicated to install and maintain. Fortunately, Airflow can meet the same needs with Kedro pipeline deployment. There is an official kedro-airflow plugin, but it doesn’t support running in Docker containers inside a Kubernetes cluster which is our preferred, most universal method.

Therefore, based on the experience of developing kedro-kubeflow, we created another plugin that we called kedro-airflow-k8s. It has the same capabilities and even the same CLI syntax as its older brother, but compiles the Kedro pipelines to Airflow DAG and deploys it by copying the file to the shared bucket which Airflow uses to synchronize Dag Bag.

machine-learning-pipelines-kedro-airflow-plugin
Machine Learning pipelines: kedro-airflow plugin

Try the plugins and let us know your thoughts!

If you’re using Kubeflow, feel free to check quickstart and the rest of the documentation. If you’re using Airflow, we have quickstart as well. If you decide to give them a try - we’re waiting for your feedback! The plugins are in Beta phase, but the main API (the way you call it from Kedro CLI) is now stable, so don’t be afraid to integrate it into your CI/CD pipelines, as we did recently.

If you want to know more please check our Machine Learning Platform and do not hesitate to contact us.

big data
technology
kubernetes
machine learning
Airflow
28 April 2021

Want more? Check our articles

getindata google data studio bigquery usage costs
Tutorial

Google Data Studio on BigQuery - usage and cost control

Data Studio is a reporting tool that comes along with other Google Cloud Platform products to bring out a simple yet reliable BI platform. There are…

Read more
geospatial analytics hadoop
Use-cases/Project

Geospatial analytics on Hadoop

A few months ago I was working on a project with a lot of geospatial data. Data was stored in HDFS, easily accessible through Hive. One of the tasks…

Read more
mamava getindata cloud google bigquery prostooleh
Success Stories

Success story: Breastfeeding supported with modern IoT and app features

Outstanding customer experience is usually backed by robust data analytics. Same applies to Mamava, a business that celebrates and supports…

Read more
highly available airflow cluster aws notext
Tutorial

Highly available Airflow cluster in Amazon AWS

These days, companies getting into Big Data are granted to compose their set of technologies from a huge variety of available solutions. Even though…

Read more
getindata grafana loki monitoring
Use-cases/Project

Why are log analytics so important in a monitoring system?

A monitoring system is a necessary component of any data platform. We can find a lot of different services that use different approaches to the same…

Read more
getindata cover nifi ingestion kafka poc notext
Tutorial

NiFi Ingestion Blog Series. PART V - It’s fast and easy, what could possibly go wrong - one year history of certain nifi flow

Apache NiFi, a big data processing engine with graphical WebUI, was created to give non-programmers the ability to swiftly and codelessly create data…

Read more

Contact us

Fill out this simple form. Our team will contact you promptly to discuss the next steps.

hello@getindata.comFist bump illustration

Any questions?

Choose one
By submitting this form, you agree to our  Terms & Conditions