Scale training across nodes and diagnose failed launches.
Recent revisions are kept. Restoring one creates a new revision.
