The PBJ Workbench features the classic workbench UI backed by the open-source Jupyter protocol pre-packaged in a runtime image. Users can easily choose this runtime image when launching a session. The open-source Jupyter infrastructure eliminates the dependency on proprietary CML code when building a Docker image, allowing runtime images to be built more quickly. The PBJ Workbench enables you to construct runtime images on Ubuntu base images (including non-Cloudera base images) and use them with the CML workbench.
ML Runtimes have been open sourced and are available in the cloudera/ml-runtimes GitHub repository. If you need to understand your Runtime environments fully or want to build a new Runtime from scratch, you can access the Dockerfiles that were used to build the ML Runtime container images in this repository.
The PBJ Workbench is available by default, but you have to select it when you launch a session.
- Click New Session.
- In PBJ Workbench , select
- Click Start Session.
Now you can use the PBJ Workbench as you would the normal workbench.
The requirements and preparatory steps for creating a PBJ Workbench are described in each section below.
PBJ Workbench setup: Python installation
PBJ Runtimes need to have Python installed, even if the Runtime is intended to run another kernel in CML, e.g. R. The minimum Python version supported is 3.7. Python can be installed from the base image’s package manager or compiled by the user.
The minimal requirements that must be satisfied by a custom PBJ Runtime image include:
- The actual Python binary or a symlink to it must exist at the following path: /usr/local/bin/python3
- The Python binary must be on the PATH under the name "python", meaning that executing the command "python" in a terminal shall start up Python.
- Executing “python --version” should result in a version greater than 3.7
- If the Runtime is configured to run a Python kernel in CML, the commands “python” and “/usr/local/bin/python3” must start up the same Python process that is registered as a Jupyter Kernel (see below).
If the method you chose to install Python does not place the Python binary under /usr/local/bin/python3, or does not create the command "python", please create appropriate symlinks.
Install Jupyter dependencies and register your kernel
First, the Jupyter Kernel Gateway, version 2.5.1 must be installed into the Docker image. This example command may need to be modified depending on the filename and path of the pip executable in the image.
RUN pip3 install "jupyter-kernel-gateway==2.5.1"
The path to the jupyter executable installed by pip should be noted, and the command to run Jupyter Kernel Gateway must be incorporated into the ML_RUNTIME_JUPYTER_KERNEL_GATEWAY_CMD environment variable in the Docker image:
ENV ML_RUNTIME_JUPYTER_KERNEL_GATEWAY_CMD="/path/to/jupyter kernelgateway"
When launching the Runtime in CML, the correct IP address - port configuration for Jupyter Kernel Gateway is set automatically by CML.
Next, a Jupyter kernel has to be registered. Each instance of the PBJ Workbench communicates with the Jupyter kernel installed in the Runtime via the Jupyter protocol. Kernels are available for a wide variety of languages and versions. Install the kernel of your choice to the image by following its installation instructions. A kernel named “python3” is registered by default when installing jupyter-kernel-gateway via pip. Installed Jupyter kernels can be listed by running the following command in a container created from the image:
path/to/jupyter kernelspec list
The name of your chosen kernel must be incorporated into the ML_RUNTIME_JUPYTER_KERNEL_NAME environment variable in the Docker image. For example, if your kernel’s name is python3, the following must be included in the Dockerfile:
Add the cdsw user
User code will be run in the image under user and group 8536:8536 . Associate these id’s with the name cdsw in the image by adding the following command to the dockerfile:
RUN groupadd --gid 8536 cdsw && \ useradd -c "CDSW User" --uid 8536 -g cdsw -m -s /bin/bash cdsw
Relax permissions so that Cloudera client config can be written
All code in the runtime container, including initial setup, will be run as the cdsw user. The initial setup includes linking client files for Cloudera data services out to their standard paths. To facilitate this, permissions to the following paths should be set so that user 8536 can write to them and to their subfolders:
Also set the permissions of the following folders and all their subfolders to 777.
- ML_RUNTIME_METADATA_VERSION environment variable and the corresponding Docker label must be set to 2.
- ML_RUNTIME_EDITOR environment variable and the corresponding Docker label must be set to "PBJ Workbench". Custom editors are not supported with PBJ Runtimes.
- Base image must be Ubuntu.
- Bash must be installed and must be configured as the default terminal used by the csdw user.
- In case the PBJ Runtime is running the R kernel, the kernel must be registered with the IRkernel package.
- The executable that is registered as a Jupyter kernel must be on PATH, must be found by the
"which" command and must be named after the programming language of the kernel. E.g. the name
of the executable must be:
- `python` in case of a Python kernel (python3 is not sufficient)
- `R` in case of an R kernel, etc.
- In the case of using a virtual / conda environment and a Python kernel, we recommend configuring PATH such that the default "pip" command corresponds to the python executable registered as a Jupyter kernel.
- CML mounts the project’s filesystem under the path /home/cdsw and erases any file installed to that path in the Runtime image. Therefore, custom Runtime images should not install anything under the home folder of the cdsw user.
- On the other hand, once the Runtime image starts up in CML, the kernel must be configured to install new packages to user site libraries under /home/cdsw. That way, newly installed packages will persist in the project’s filesystem.
- The package
xz-utilsmust be installed on the Runtime image.