Flattened expressions - so that AND never contains a child that is AND, and OR never contains a child that is OR - is a canonized form that seems to have more advantages than disadvantages. It's never larger than the original (unlike CNF and DNF), frequently smaller than the original, doesn't take much effort to canonize, and allows the multiple consumers of expressions (simplifications, planner rules) to make some assumptions that yield good results efficiently.
If you are only producing expressions you want Calcite to be liberal in what it accepts; but if you are consuming expressions (e.g. maintaining a rule) you want Calcite to be conservative in what it accepts. You can't please both producers and consumers - it's a zero sum game - but flattened expressions seemed to be a reasonable compromise. I hope there's a simple process to get your expressions flattened. If there's not, log a bug, so that there's a documented process. By the way, your note reminds me that we're not as diligent at flattening OR as we are at flattening AND. We should fix that. Julian On Wed, Aug 9, 2023 at 9:09 AM Ian Bertolacci <ian.bertola...@workday.com.invalid> wrote: > > Hello, > I’m curious about why Filter requires that the condition be a flattened tree? > We have some transformations which occasionally result in non-flat trees, > which causes issues for us in testing. > > I know we can avoid this by constructing expressions using RelBuilder or the > RexUtil.composeConjunction, but our transformers work very similar to > RexShuttle, which simply clones a RexCall on reconstruction. > So there are situations where a subexpression to an `AND` or `OR` expression > result in a nested `AND` or `OR` expression. > One solution is to use RexUtil.flatten, but that only applies flattening for > subexpressions with the same operator, so that expressions like `AND( A, AND( > B, C ), OR( E, OR( F, G ) )` only get partially flattened to only gets > flattened to `AND( A, B, C, OR( E, OR( F, G) )` which still isn’t flat; the > correct flattening would be `AND( A, B, C, OR( E, F, G ) )` > Similarly, something like `AND( A, AND( B, C ) ) == …` wouldn’t get flattened > at all. > > Is there a real reason for having this assertion? > Would removing it cause problems? > Thanks! > -Ian J. Bertolacci
