> PlanState.has_recursivescan seems like a complete kluge. Can't it just be
> removed? It looks to me like it is working around bugs that hopefully aren't
> there anymore. There is certainly no reason why a recursive CTE should be
> more in need of rescanning than any other kind of plan.
I don't think so. Recursion plan needs the hash table used by sublan
be re-created at each recursion loop stage. Remember that in each
evaluation of recursive plan, the recursive name is replaced by a
working table which is holding previous evalution result of recursion
stage. Thus the hash table corresponding to the work table needs to
be re-created.
> If it is needed then
> the current implementation is completely broken anyway, since it would only
> detect a RecursiveScan node that is directly underneath an agg or hash node.
Yeah, that's right. What I have in my mind is to implement something
similar to UpdateChangedParamSet family like mechanism which will
inherit working table change event to child node.
--
Tatsuo Ishii
SRA OSS, Inc. Japan