Distributed data computing in TO RUN clause.

[brightcoder01]

In SQLFlow TO RUN clause, it will call a python function to do the data processing/computing. Such as use TSFresh to extract features from time series data.

• If the data size is small, the python function can be executed in a single process.
• If the data size is large scale, we need execute the python data computing in a distributed way.

There are two options for large-scale data computing with python:

• Dask
  MaxCompute Support: ×
  Kubernetes Support: √ link
  TSFresh integration: The official distributed computing support for TSFresh is Dask. link

• Mars
  MaxCompute Support: √
  Kubernetes Support: √ link - Verifying using Minikube, downloading image is too slow. We can build the image on our Mac at the first step. link
  TSFresh integration: no pre-made solution, need development from mars team, issue.

The compare between Dask and mars: issue

• Ray?

[wangkuiyi]

Thanks to @brightcoder01 for raising this question.

I reorganized the comparison in the above comment into a table.

	Dask	Mars
Kubernetes	Yes	No
MaxCompute	No	Yes

It seems that the question is

• When SQLFlow users write SELECT ... TO RUN a_python_func, s/he might want the Python function executed in parallel to process large datasets. Should SQLFlow take care of the parallel execution use Mars/Dask/Ray/etc, or should the author of a_python_func take care of the execution by calling Mars/Dask/Kubernetes API to file a parallel job?

I think the answer is to let the author of a_python_func to make the choice among Mars, Dask, Ray, Kubernetes API, etc.

[brightcoder01]

Add notes from discussion:

• The python function can be stored in the form of Docker image in the Model zoo.