Proposal: Support saving the model with PMML format

[weiguoz]

Currently, We train a model, which is used for prediction later. This works fine under SQLFlow.
But, when we trained a model by SQLFlow, then use the model for prediction in other platforms/tools, we will encounter compatibility problems. Meanwhile, it's not easy to make such other platforms support SQLFlow format.

So, I propose SQLFlow support exporting different types of model formats, e.g. PMML which is widely used in the industry. Generally, it will be more flexible in a user-defined way to specify the model format.

Proposals:

1. Specify the model format by a suffix of the model name in the SQL, like:

SELECT *
TO TRAIN DNNClassifier
COLUMN col1, col2, col3, col4
INTO my_model.pmml   --- my_mode.{user-defined model format}

2. Specify the format in the SQLFlow server level via an environment variable.
   This method lacks flexibility but has no addition to SQL expressions.

I like the first proposal because of flexibility.
How do you think? @Yancey1989 @typhoonzero @tonyyang-svail @wangkuiyi @llxxxll @shendiaomo @lhw362950217

[shendiaomo]

I prefer the first proposal's flexibility. However, before we start to dig into details, I think we should first estimate carefully whether PMML is necessary. For example, will TensorFlow-serving be more appropriate for such requirements and/or easier to develop?

[tonyyang-svail]

For the first proposal, do we need to deal with INTO database_name.table_name with database_name and table_name accidentally being my_model and pmml?