Tutorial | Getting started with CI/CD pipelines with Dataiku#

Before starting on your specific use case, begin here to understand the requirements and architecture and obtain the necessary assets.


As an overview, we recommend reading our blog post on Continuous integration and continuous deployment (CI/CD) in Dataiku that describes the concepts and questions in order to fully optimize CI/CD.

For you to be at ease with the content, you need to know about Dataiku flows and automation scenarios. We strongly recommend following the Operationalization series from the Dataiku Academy.

In addition, a good understanding of CI/CD orchestration, and some Python skills (including pytest ideally) are needed.

In this article, we will describe the common logic and the use case. Once familiar with it, you can proceed to the setup that is closest to your need:

  • Using Jenkins to deploy an API endpoint with API Deployer.

  • Using Jenkins to deploy an analytics flow with Project Deployer.

  • Using Azure pipelines to deploy an analytics flow with Project Deployer.

  • Using Jenkins to deploy an analytics flow directly (i.e. without Project Deployer).

You’ll find the links to these articles at the bottom of the page.


Operationalization requires a minimal setup of your platforms in order to leverage the different types of Dataiku nodes:

  • The Dataiku Design node is your development environment and is where you can design your Flow and build or improve your data logic. Since we’ll be designing the project and the API endpoints on a Design node, you’ll need to have one available. For more information, visit Installing and setting up Dataiku.

  • Automation nodes are made to run production analytics flows. You will find the installation guide in the documentation. If you are looking to deploy an analytics flow, you will need two of those. In reality, you only need one, but in our pipeline, we will leverage a pre-prod instance to run some tests before moving to a production one.

  • API nodes are made to serve live prediction requests. You can either deploy a single static API node or a kubernetes cluster of API nodes. In the examples related to API endpoints deployment, we need two static nodes, one for testing and one for production. See the installation guide in the reference documentation.

In the API node configuration, you will need to set up the API Deployer. You can see how to do that in the reference documentation.


In the Automation node configuration, if you want to use the Project Deployer, you will also need to configure it as well, as explained in the setup documentation.


In addition to these Dataiku requirements, you need an orchestrator with the ability to execute Python code. Our examples are using mostly Jenkins, but there is one on also on Azure Pipelines. Generally speaking, any pipeline tool will work, as long as you can execute Python code with it. (Technically, you can also use Dataiku HTTP REST APIs, but this will be much more complex so we strongly advise to go for the Python client).

We will also mention the usage of an Artifact repository, where we will archive the project file before moving it to production. We showcase using Artifactory, but any tool that you can push a zip file through the orchestrator will work as easily (such as Azure devops Artifacts, Maven, Nexus, or a file system such as AWS S3 or Google Cloud Storage)

Lastly, the code of our pipeline itself will be stored in a code repository. In the examples, we store it in GitHub, but this can be any code repository compatible with your orchestrator.



Automation and API nodes require a specific commercial license. If you are a customer and want to know more about those features, please contact your Customer Success Manager to see available options.

The project#

We will use the Prediction churn sample project. You can directly deploy it on your Dataiku instance.

  1. From the Dataiku homepage, click +New Project > Sample projects > Basics > Predicting churn.


You can also download the starter project from this website and import it as a zip file.


For the Analytics flow use case, you need to add a twist at the end with a Python recipe. The code itself is not really important. We just need it to showcase code analysis. You can even leave the default code generated at the recipe’s creation.

This is not required for the real-time prediction use cases where you will just need the model.


Lastly, make sure the Flow is working properly by building the output dataset. You may need to retrain and re-deploy the saved models depending on your configuration.

Pipeline description#

Our pipeline will always contain the sames stages:



Get our pipeline environment ready. Usually, we will set up global variables, ensure the orchestration workspace is able to run our code, fetch the pipeline code, etc.


This stage checks that the project we want to operationalize is “good enough”. This greatly varies from one organization to another and you can check naming conventions, complexity rules, documentation, or anything that would prevent the project to be ready for production. This usually relies on specific code to write to fetch the data and control it.


This stage build the consistent package of the artifact we want to move to production (a bundle in the case of an analytics flow, an API package in the case of a real-time prediction). This is also where we will archive this artifact for auditing/history purposes. This stage is simple but critical as this is where you need to make sure of what you are actually moving to production. Although there may be alternatives, we strongly suggest to use the bundle & API package artifacts as they are made exactly for such a use case.


At this stage, the artifact will be deployed to a test instance. Then all tests will be run and checked for errors. In the case of errors, the pipeline will stop, otherwise, the move to production will occur. Note that you can stop your pipeline at this stage: instead of Continuous Deployment, you will then achieve Continuous Delivery. Up to you to manage the deployment after then (it can be through another tool or manually).


This is the final stage where the artifact will be deployed to the production instance. We will limit the test to basic ones and incorporate a rollback mechanism.

What’s next?#

Now, you have the big picture! Let’s dive in the specific use case you want to reproduce.