Allow pipeline to create pipeline

[MatthieuBizien]

Like #3549 say, "A run of a pipeline is like a function call". Based on that analogy, we can't right now call a function from another function: f(g(x)) doesn't make any sense for the current dvc model. As discussed in #331, it would allow to use very flexible workflows by allowing to dynamically created pipelines.

Some potential applications:

• Incremental processing of dynamically generated files Incremental processing or streaming in micro-batches #331
• Reconfigurable pipelines Reconfigurable pipelines #1462
• Build matrix dvc: consider introducing build matrix #1018
• Hyperparameters tuning ML experiments and hyperparameters tuning #2799

A possible interface could be:

• dvc run could specify a yaml file to use, instead of dvc.yml
• We allow that file to the output of another pipeline step
• Then dvc run -n "create_pipe" -o generated_pipe.yml generate_pipe.sh would create the generated_pipe.yml file, and dvc repro -c generated_pipe.yml would execute all the steps dynamically defined.

[skshetry]

Without the generation overhead, we could introduce depends_on key in dvc.yaml, but it might not be straightforward,
as we create a pipeline based on relationship of outputs-dependencies which might easily create cyclic graph.

stages:
   stage1:
     ...
   stage2:
    depends_on: stage1
    ...

[dmpetrov]

The first implementation is coming #4734. The next step is #331.

Let's close this issue and move all the discussions to "umbrella" issues #3633 & #331.

[MatthieuBizien]

Thanks for #4734, it address one of the major limitation of DVC! But this issue is about something different, more powerful, so I don't think it should be closed.

[efiop]

@MatthieuBizien Have you considered the alternative approach that @skshetry suggested? For now we are not sure if we will ever allow dynamic generation like you've suggested, as there are alternatives like parametrization or generating the whole pipeline programmatically.