Four models, one job each: what our own logs say about splitting the work