It's such a comprehensive explanation, Thanks Jacques, I get it. Jacques Nadeau <[email protected]> 于2021年11月27日周六 上午1:10写道：
> > > > Yes, I'm referring to the RelNode's row type, the planBefore and > > planAfter's row types are different in field name. > > > > You cannot rely on field names from individual RelNodes. It's just the way > Calcite was designed. > > I'm trying to convert the optimized RelNodes tree to a sql with same > > semantics with the original RelNodes, > > If the field name is missing, the sql will be incorrect. > > > > You should be using the RelRoot's validated row type to generate the field > names of the output, not the RelNode's. They will be correct. If you're > constructing the plan by hand, you'll need to create a rowtype independent > of the relnodes to maintain the field names through the process. > > > > Thanks JiaTao and Jing, I understand the fact that alias is trivial in > > rules, But how can Calcite ensure the physical plan > > which transformed from optimized logical plan outputs the correct column > > names? > > > > By using the RelRoot's validated row type. >
