ClearML pipelines can be build from tasks, functions, and decorated functions, according to the examples in None . I am guessing there was a certain way of doing things, that allgegro.ai started with and then there was a pain that the other approaches solved. So, what are the rules-of-thumb when to use which approach?