Debugging Issues with Experiments

This topic lists some common issues to watch out for during an experiment's build and execution process:

Experiment spends too long in Scheduling/Built stage

If your experiments are spending too long in any particular stage, check the resource consumption statistics for the cluster. When the cluster starts to run out of resources, often experiments (and other entities like jobs, models) will spend too long in the queue before they can be executed.

Resource consumption by experiments (and jobs, sessions) can be tracked by site administrators on the Admin > Activity page.

Experiment fails in the Build stage

During the build stage Cloudera Machine Learning creates a new Docker image for the experiment. You can track progress for this stage on each experiment's Build page. The build logs on this page should help point you in the right direction.

Common issues that might cause failures at this stage include:
  • Lack of execute permissions on the build script itself.
  • Inability to reach the Python package index or R mirror when installing packages.
  • Typo in the name of the build script (cdsw-build.sh). Note that the build process will only execute a script called cdsw-build.sh; not any other bash scripts from your project.
  • Using pip3 to install packages in cdsw-build.sh, but selecting a Python 2 kernel when you actually launch the experiment. Or vice versa.

Experiment fails in the Execute stage

Each experiment includes a Session page where you can track the output of the experiment as it executes. This is similar to the output you would see if you test the experiment in the workbench console. Any runtime errors will display on the Session page just as they would in an interactive session.