Switch branches/tags
Find file History


Jupyter Notebook

This folder contains the initialization action jupyter.sh to quickly setup and launch Jupyter Notebook, (the successor of IPython notebook) and a script to be run on the user's local machine to access the Jupyter notebook server.

Note: This init action uses Conda and Python 3. Python 2 and pip uesrs should consider using the jupyter2 action.

Using this initialization action

You can use this initialization action to create a new Dataproc cluster with Jupyter installed:

  1. Use the gcloud command to create a new cluster with this initialization action. The following command will create a new cluster named <CLUSTER_NAME>.

    # Simple one-liner; just use all default settings for your cluster. Jupyter will run on port 8123
    # of your master node.
    gcloud dataproc clusters create my-dataproc-cluster \
        --initialization-actions \
            gs://dataproc-initialization-actions/jupyter/jupyter.sh \
  2. Run ./launch-jupyter-interface to connect to the Jupyter notebook running on the master node. This creates a SOCKS5 proxy to the master node and launches a Google Chrome window that uses this proxy. Note: you will need to edit the script to point it at the Chrome installation path for your operating system. Alternatively, follow the instructions in connecting to cluster web interfaces.


There are various options for customizing your Jupyter installation. These can be provided as metadata keys using --metadata.

  • JUPYTER_PORT=: Port on which the Jupyter server runs
  • JUPYTER_CONDA_PACKAGES=: List of Conda packages to install.
  • INIT_ACTIONS_REPO=: Repo to clone to find other scripts to install/configure Conda and Jupyter.

For example to specify a different port and specify additional packages to install:

gcloud dataproc clusters create my-dataproc-cluster \
    --metadata "JUPYTER_PORT=8124,JUPYTER_CONDA_PACKAGES=numpy:pandas:scikit-learn" \
    --initialization-actions \
        gs://dataproc-initialization-actions/jupyter/jupyter.sh \
    --properties spark:spark.executorEnv.PYTHONHASHSEED=0,spark:spark.yarn.am.memory=1024m \
    --worker-machine-type=n1-standard-4 \

Notebooks are stored and retrieved from the cluster staging bucket (Google Cloud Storage) at gs://<staging-bucket>/notebooks/. By default, clusters in your project in the same region use the same bucket. You can explicitly provide --bucket=gs://<some-bucket> to the same value to share notebooks between them.

Internal details


jupyter.sh handles configuring and running Jupyter on the Dataproc master node by doing the following:

  • clones the dataproc-initialization-actions git repo/branch specified in the INIT_ACTIONS_REPO and INIT_ACTIONS_BRANCH metadata keys
    • if these two metadata keys are not set during cluster creation, the default values https://github.com/GoogleCloudPlatform/dataproc-initialization-actions.git and master are used
    • this is provided so that a fork/branch of the main repo can easily be used, eg, during development
  • executes conda/bootstrap-conda.sh from said repo/branch to ensure miniconda is available
  • executes jupyter/internal/setup-jupyter-kernel.sh and jupyter/internal/launch-jupyter-kernel.sh from said repo/branch
    • configures jupyter to use the PySpark kernel found at jupyter/kernels/pyspark/kernel.json
    • configures jupyter to listen on the port specified by the metadata key JUPYTER_PORT, with a default value of 8123
    • configures jupyter to use auth token JUPYTER_AUTH_TOKEN, with a default of none.
    • loads and saves notebooks to gs://$DATAPROC_BUCKET/notebooks/, where $DATAPROC_BUCKET is the value stored in the metadata key dataproc-bucket (set by default upon cluster creation and overridable). Note that all clusters sharing a $DATAPROC_BUCKET will share notebooks.
    • launches the jupyter notebook process

NOTE: to be run as an init action.


launch-jupyter-interface.sh launches a web interface to connect to Jupyter notebook process running on master node.

  • sets a path for the local OS to the Chrome executable
  • setup an ssh tunnel and socks proxy to the master node
  • launch a Chrome instance that uses this ssh tunnel and references the Jupyter port.

NOTE: to be run from a local machine

Important notes

  • This initialization action clones this repo at master to run other scripts in the repo. If you plan to copy jupyter.sh to your own GCS bucket, you will also need to fork this repository and specify the metadata keys INIT_ACTIONS_REPO and INIT_ACTIONS_BRANCH.
  • This initialization action runs the conda init action, which supports the metadata keys CONDA_PACKAGES and PIP_PACKAGES. You can also use these to install additional packages.