Tutorial | Code notebooks and recipes#

Get started#

Dataiku not only has natively-integrated Jupyter notebooks for experimentation, but also has a way to deploy them as recipes into production workflows.


In this tutorial, you will:

  • Create, edit, publish, and unload Jupyter notebooks.

  • Create a code recipe from a notebook.

  • Sync code back and forth between a notebook and a recipe.

  • Run a code recipe in the Flow.

See also

This tutorial only covers the use of Python notebooks and recipes in Dataiku, but the workflow is similar for other languages. See the following for resources on R and SQL.


  • Access to an instance of Dataiku.

  • Some familiarity with coding in Python and using Jupyter notebooks.

  • The requested permissions for code execution.

Create the project#

The first step is to create a new Dataiku Project. We will work with a sample project containing data from the fictional Haiku T-Shirt company.

  1. From the Dataiku Design homepage, click + New Project > DSS tutorials > Developer > Code Notebooks & Recipes.

  2. From the project homepage, click Go to Flow (or g + f).


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

Use a code notebook#

Given their usefulness for data science, Jupyter notebooks are natively embedded in Dataiku and tightly integrated with other components, which makes them easy to use in various ways.

Create a Jupyter notebook#

Depending on your objectives, you can create a Jupyter notebook in Dataiku in a number of different ways. In this exercise, we will create a notebook from a dataset, which simplifies reading in the dataset of interest using the Dataiku API.

  1. From the Flow, select the orders dataset.

  2. In the right panel, select the Lab menu (with the microscope icon).

  3. In the Code Notebooks section, click New.

  4. From the notebook options, select Python.

  5. Name the notebook orders analysis.

  6. Click Create, leaving the default option to read the dataset in memory using Pandas.

Dataiku screenshot of dialog for creating a Python notebook.

Inspect a notebook’s starter code#

The newly created notebook contains some useful starter code:

  • The first cell uses the built-in magic commands to import the numpy and matplotlib packages.

  • The second cell imports other useful packages, including dataiku.

  • The third cell reads in the orders dataset and converts it to a Pandas dataframe.

  • The fourth cell contains a function that performs some basic analysis on the columns of the dataset.

Dataiku screenshot of starter code in a Python notebook.


The starter code of a notebook created from a dataset will have already read in the chosen dataset to a df variable, whether it may be a Pandas, R, or Scala dataframe.

Edit code in a notebook#

You can edit the starter code as well as write your own code in the same way you would outside of Dataiku.

In this very simple exercise, we will slightly modify the existing starter code:

  1. Delete limit=100000 from the second line of code in the third cell to remove the default dataset sampling. After removing it, the line of code should look like this:

    df = dataset_orders.get_dataframe()
  2. Type df.head() right under the one above. The code in the third cell should now look like this:

    # Read the dataset as a Pandas dataframe in memory
    # Note: here, we only read the first 100K rows. Other sampling options are available
    dataset_orders = dataiku.Dataset("orders")
    df = dataset_orders.get_dataframe()
  3. Run the first three cells to read in the orders dataset and display the first five rows of the dataset.

  4. Run the fourth and last cell (pdu.audit(df)), which is part of the starter code, to display some basic information about the columns of the orders dataset.

  5. Click the Save button (or use the shortcut Ctrl + S / Cmd + S for Mac) to save your progress.

Dataiku screenshot of an edited Python notebook.


In addition to datasets, it’s also possible to create Jupyter notebooks from machine learning models. For more information, consult the reference documentation.

Publish a notebook to a dashboard#

For a collaborative platform like Dataiku, the ability to share work and analyses is of high importance. Dataiku allows you to save static exports (non-interactive snapshots) of Jupyter notebooks in an HTML format, which can be shared on dashboards.

To share the notebook on a dashboard:

  1. Open the Actions tab of the notebook.

  2. Click Publish > Dashboard.

  3. Click Create to create the notebook as an insight and publish that insight to the first slide of the project’s default dashboard.

  4. In the dashboard, expand the notebook tile by dragging the corners of the insight.

  5. In the Tile settings, click Show code to include the code and not just the printed output.

  6. Click Save and then View to see how the notebook insight appears on the dashboard.

Dataiku screenshot of a code notebook insight in a dashboard from the Edit tab.


Publishing a static snapshot of a notebook to a dashboard also adds it to the list of saved insights. To learn more about sharing Jupyter notebooks as insights, see the reference documentation.

Unload a notebook#

Finally, once you’re done working in a Jupyter notebook for the time being, you can optimize its computational efficiency by killing the kernel. To do this:

  1. Navigate to the Notebooks page (g + n).

  2. Check the box to select the orders analysis notebook.

  3. In the right panel, in the Actions tab, click Unload to kill the kernel.

Use a code recipe#

Code notebooks allow for free experimentation, but you’ll need code recipes to build outputs in your Flow.

Create a recipe from a notebook#

Code recipes can be created from scratch directly from the Flow. However, you can also create them from existing notebooks. This can be particularly useful for deploying exploratory work from notebooks to the Flow.

Let’s create a Python code recipe from the Jupyter notebook.

  1. Navigate to the Notebooks page (g + n), and open the orders analysis Python notebook.

  2. Within the notebook’s header, click + Create Recipe.

  3. Verify Python recipe is selected, and click OK to confirm.

  4. Verify the orders dataset is selected as the input (since that is the dataset used in the notebook).

  5. Under Outputs, click + Add. Name it orders_by_customer.

  6. Click Create Dataset.

  7. Click Create Recipe.

Dataiku screenshot of the dialog for creating a recipe from a notebook.

Edit code in a recipe#

In the resulting recipe, all the code from the Jupyter notebook has been transferred to the recipe code editor. Notice that Dataiku has added a number of commented out lines, each of which shows the beginning of a notebook cell. This way, if we need to edit the recipe in a notebook again, our existing cells are maintained.

The editor has also added two lines for the recipe output based on the name of the output dataset we created in the recipe dialog. We’ll discuss this below.

Now let’s group the orders data by unique customers. Although we could accomplish this with a visual Group recipe, it can also be done with code.

  1. Comment out the following lines of code:

    • df.head()

    • pdu.audit(df)

  2. In a new line below df = dataset_orders.get_dataframe(), copy-paste the following code:

    orders_by_customer_df = df.assign(total=df.tshirt_price*df.tshirt_quantity


This code block creates a new dataframe with rows grouped by customer_id. For each customer, we’ve computed the average number of pages on the Haiku T-shirt website visited by the customer during orders, and the sum total of the value of orders made by the customer, where the value of each order is the price of each t-shirt multiplied by the number of t-shirts purchased.

Finally, Dataiku has added lines for the recipe output. However, it cannot know which dataframe (df or orders_by_customer_df) we want to output as the orders_by_customer dataset. Accordingly:

  1. In the last line of code, change pandas_dataframe to orders_by_customer_df.

  2. Click Validate to check the validity of the code. It should display Validation successful.

  3. Click Run (or type @ + r + u + n) to execute the recipe, and then explore the output dataset.

Dataiku screenshot of a Python recipe having been run.

Iterate between a recipe and a notebook#

Notice that the output dataset orders_by_customer does not contain the customer_id column, even though this was the key we grouped by. We’d like to have it for reference.

Let’s prototype the revised code in the notebook attached to the recipe.

  1. Reopen the Python recipe (clicking Parent Recipe is one option).

  2. In the header, click Edit in Notebook.

  3. Click Override Notebook Content.


Previously we edited the recipe without saving the changes back to the recipe, creating this mismatch. Overriding the notebook content re-syncs the recipe and notebook.

Now we can interactively test the recipe code in a notebook.

  1. Uncomment the df.head() line and change df to orders_by_customer_df, so that the new line is as follows:

  2. Run the first three cells. The output shows that the orders_by_customer_df dataframe has the customer_id information; however, the dataframe has a hierarchical index.

  3. In order to flatten the index, add .reset_index() to the code that defines the dataframe so that it looks like the following:

    orders_by_customer_df = df.assign(total=df.tshirt_price*df.tshirt_quantity
  4. Re-run the third cell to see how the dataframe has changed.

  5. In the header, click Save Back to Recipe.

Dataiku screenshot of a Python notebook.
  1. Comment out orders_by_customer_df.head(), Validate, and Run the recipe again.

Now the output dataset contains a customer_id column.

Dataiku screenshot of a dataset after a Python recipe.

What’s next?#

This tutorial used the built-in code environment, but often you’ll want to use your own. Learn about code environments in How-to | Create a code environment!

See also

To learn more, see the reference documentation on Code notebooks and Recipes based on code.