The ML Runtimes 2022.04-1 Release includes a technical preview version of the new workbench architecture, the PBJ (Powered by Jupyter) Workbench. In the previous Workbench editor, the code and output shown in the console (the right-hand pane in the image below) were passed to and from Python, R, or Scala via a Cloudera-specific, custom messaging protocol. In the PBJ Workbench, on the other hand, the code and output are now passed to and from the target language runtime via the Jupyter messaging protocol. They are handled inside the runtime container by a Jupyter kernel and rendered in your browser by JupyterLab’s client code.
This may seem like a subtle change, but it will provide CML users with some major benefits. First, the behavior of user code and third-party libraries on CML will be more consistent with its behavior in Jupyter-based environments. That means that a wider variety of rich visualization libraries will work out of the box, and in cases where rich visualization libraries do not work, error messages in the CML console and the browser console will be easier to google. Likewise, dependency conflicts between kernel code and user code will be rarer, and when they do occur they will be easier for customers to diagnose and fix. To give you a taste of what this higher degree of consistency is like, note that Python 3’s input() function now works. Go ahead and try it out!
Second, customers will no longer need to build runtime images starting from Cloudera base images and will no longer need to restrict themselves to languages and versions that Cloudera has packaged. Any combination of base image, target language, and language version can be used with the PBJ Workbench as long as a Jupyter kernel is available for that combination.
You can try it out by running a PBJ Workbench Python session using a CML April or newer Workspace. The look and feel of the workbench will be more or less unchanged. Under the hood, however, the way that code and outputs are rendered and passed between the web app and the Python interpreter have been re-engineered to better align with the Jupyter ecosystem.
The Technical Preview documentation is available here.
