Wu Yi

Some background: https://github.com/sql-machine-learning/sqlflow/issues/2996

How about `TO RUN a_docker_registry/sqlflow-run:v1/binning.py`

Generally, SSA is generated from AST, not from IR, and the AST should have variables and assignment statements. Yet the SQL language often used without variables and assignments.

We need to output both workflow and compiled step like: 1. `sqlflow -workflow a.sql -o a.yaml` compile a SQL program to a workflow. 2. `sqlflow -step a.sql -o a.py` compiles...

No, both workflow and step is needed. If we change each step to a pure Python code for execution, the command `sqlflow -workflow a.sql -o a.yaml` contains everything. Yet when...

Can we write libtorch dataset transform functions to achieve this?

Maybe the sqlflowserver version have very little chance to interference with the workflow step. The sqlflowserver is only generating the workflow and run, yet the version of `TO RUN` can...

Reasonable. In previous discussions with @Yancey1989 and @shendiaomo, we've been talked about this. Yet, We can somehow derive the feature columns in compile-time. ```sql CREATE TABLE a (...); INSERT INTO...

- Support mathematical programming using SQL

尽量使用多进程方式。v2 python端inference比较难控制GPU初始化或多stream。使用多进程模型，可以指定多个进程共用一个GPU也可以指定多个进程分别使用多个GPU。 另建议升级到最新版fluid，inference会有更多选项。