> The equal() updates I installed yesterday (to fix the "don't know
> whether nodes of type 600 are equal" problem) have had an unintended
> side effect.
> Am I right in thinking that UNION (without ALL) is defined to do a
> DISTINCT on its result, so that duplicates are removed even if the
> duplicates both came from the same source table? That's what 6.4.2
> does, but I do not know if it's strictly kosher according to the SQL
> spec.
Yup, that's the way it should be...
> If so, the code is now busted, because with the equal() extension in
> place, cnfify() is able to recognize and remove duplicate select
> clauses. That is, "SELECT xxx UNION SELECT xxx" will be folded to
> just "SELECT xxx" ... and that doesn't mean the same thing.
:(
> Assuming that 6.4.2 is doing the Right Thing, I see two possible
> fixes:
> (1) simplify equal() to say that two T_Query nodes are never equal, or
> (2) modify the planner so that the "select distinct" operation is
> inserted explicitly, and will thus happen even if the UNIONed selects
> are collapsed into just one.
Sounds to me like (2) is the correct way to do it, as long as the
explicit "sort/unique" happens only if the query is collapsed. I haven't
looked at this code, and have no experience with cnfify(), but at the
time it decides to collapse into a single select, can't it ensure that a
sort/unique is done instead? Or is that what you are suggesting?
Wouldn't want to have two sort/unique operations on top of each other...
- Tom
