Faster Spark jobs with less wasted memory.
Recent revisions are kept. Restoring one creates a new revision.
