Have a look at our YouTube video for a quick walkthrough of Orchest! 💪

This quickstart will follow an example explaining how to build data science pipelines in Orchest and touching upon some core principles that will be helpful when you get to building your own pipelines. The example pipeline will download the sklearn California housing dataset, explore the data, train some classifiers, and in the final step collect the results from those classifiers.


The resulting pipeline from this quickstart.

Please refer to How Orchest works to get a more in-depth explanation of concepts within Orchest.

For the impatient

As Miguel Grinberg would say: “If you are the instant gratification type, and the screenshot at the top of this article intrigued you, then head over to the Github repository for the code used in this article. Then come back to learn how everything works!”

To get started with the pipeline in Orchest you can import the GitHub repositories URL https://github.com/orchest/quickstart through the UI:


Your first project

To start, make sure you have installed Orchest and started it:

# Make sure to be in the root-level orchest directory.
./orchest start

For the quickstart we will create a new project named quickstart. After creating the project, you will see that it does not yet have any pipelines.



All code in this quickstart is written in Python, nevertheless, we do also support other languages such as R.

Get California housing data

The logical next step is to create the first pipeline called California housing and open the pipeline editor. This will automatically boot an interactive session so you can interactively edit the Python script we create (the other steps will be notebooks!):

  1. Create a new step by clicking: + new step.
  2. Enter a Title and File path, respectively Get housing data and get-data.py.
  3. Make sure to save your progress with save*.

Now we can start writing our code through the familiar JupyterLab interface, simply press edit in JupyterLab (making sure you have the step selected) and paste in the following code:

import orchest
import pandas as pd
from sklearn import datasets

# Explicitly cache the data in the "/data" directory since the
# kernel is running in a Docker container, which are stateless.
# The "/data" directory is a special directory managed by Orchest
# to allow data to be persisted and shared across pipelines and
# even projects.
print("Dowloading California housing data...")
data = datasets.fetch_california_housing(data_home="/data")

# Convert the data into a DataFrame.
df_data = pd.DataFrame(data["data"], columns=data["feature_names"])
df_target = pd.DataFrame(data["target"], columns=["MedHouseVal"])

# Output the housing data so the next steps can retrieve it.
print("Outputting converted housing data...")
orchest.output((df_data, df_target), name="data")

As you can see, we have highlighted a few lines in the code to emphasize important nuts and bolts to get a better understanding of building pipelines in Orchest. These nuts and bolts are explained below.

First we start with explaining line 11 in which we cache the data in the /data directory. This is actually the userdir/data directory (from the Orchest GitHub repository) that gets bind mounted in the respective Docker container running your code. This allows you to access the data from any pipeline, even from pipelines in different projects. Data should be stored in /data not only for sharing purposes, but also to make sure that jobs do not unnecessarily copy the data when creating the snapshot for reprodicibility reasons.

Secondly, line 19 showcases the usage of the Orchest SDK to pass data between pipeline steps. Keep in mind that calling orchest.transfer.output() multiple times will result in the data getting overwritten, in other words: only output data once per step.

To run the code, switch back to the pipeline editor, select the step and press run selected steps. After just a few seconds you should see that the step completed successfully. Let’s check the logs to confirm, the logs contain the latest STDOUT of the script.


Remember that running the code will output the converted housing data, in the next step we can now retrieve and explore that data!

Data exploration

Now that we have downloaded the data, the next pipeline step can explore it. Create another pipeline step with Title Data exploration and File path explore-data.ipynb, and connect the two pipeline steps.


You can get the code for this pipeline step from the explore-data.ipynb file in the GitHub repository.

Maybe you already noticed the imports in the previous step:

import orchest
import pandas as pd
from sklearn import datasets

These dependencies are satisfied by default, because the environments are based on the Jupyter Docker Stacks which already contains a number of common data science packages. In this data exploration step however, we make use of Vaex to showcase how environments let you install additional packages.

Go to Environments in the left pane menu and inspect the Python 3 environment. Here you can see that pip install vaex is added to the Environment set-up script.

Finalizing the pipeline

To end up with the final pipeline, please refer to the For the impatient section to import the pipeline. You can also build the pipeline from scratch yourself!


A successful pipeline run of the final pipeline.


The interactive session does not shut down automatically and thus the resources will keep running when editing another pipeline, you can shut down the session manually by clicking on the shut down button. Of course all resources are shut down when you shut down Orchest with ./orchest stop.