Build physical filter eagerly while constructing ParquetExec operator

[yjshen]

We should avoid leaking logical expression to physical operators.

https://github.com/apache/arrow-datafusion/blob/master/datafusion/src/physical_optimizer/pruning.rs#L93-L101

https://github.com/apache/arrow-datafusion/blob/master/datafusion/src/physical_plan/file_format/parquet.rs#L77-L86

[yjshen]

@alamb could you please provide some context on why we are using Expr instead of PhyscialExpr for ParquetExec in the first place?

[alamb]

@alamb could you please provide some context on why we are using Expr instead of PhyscialExpr for ParquetExec in the first place?

I don't know of any original rationale -- I think @yordan-pavlov 's initial implementation was in terms of Expr

I think it would be fine to use physical exprs instead (as we can always translate from Expr --> PhysicalExpr)

[yordan-pavlov]

@yjshen when I first introduced parquet predicate push-down for Data Fusion in apache/arrow#9064 I was focused on reusing existing code for evaluating dynamically generated expressions. I don't remember any particular reason for using logical instead of physical expressions. If you think using physical expressions (instead of logical) would make the code better, that's fine with me.

[yjshen]

Thanks @yordan-pavlov for the context! I am asking here to make sure this proposal won't break any hypothesis I'm not aware of.