Skip to main content

Recommendations

  • Stop the notebooks you do not use. A running notebook reserves its GPU even if nobody runs anything, and that reservation shows as occupied for the whole Client.

  • One lab per project, not per person. Labs are shared; naming them after the project avoids duplicates when the team changes.

  • Start with the default size. Move up only when a job is rejected; the message tells you which resource was missing.

  • Everything in /home/jovyan, and the mount path untouched. It is the only folder that survives a restart. A notebook with the volume mounted on another path works the same until the first restart, and then loses everything (why).

  • Large data on a Data Volume, not in the notebook folder. Data volumes mount in several notebooks and survive deleting one.

  • Turn into a pipeline anything you run more than twice. A notebook is for exploring; a pipeline leaves a record and can be queued.

  • Nothing confidential in pipeline names, parameters or metrics. That information is visible to other users of the pipelines platform.

  • Read Advanced configuration before going beyond the default values. It collects the real cases where a freedom the panel allows ended in data loss or in a notebook that does not start.

Next step​