On Tue, 21 Apr 2026 at 19:29, Andrei Lepikhov <[email protected]> wrote: > > On 17/04/2026 10:56, Andrei Lepikhov wrote: > > The best-known problematic code example causing this issue is > > apply_scanjoin_target_to_paths(), and the current_rel/final_rel game from > > commit > > 0927d2f46dd. Quickly fixing it, I see some more combinations have emerged: > > On closer inspection, it looks like all the detected cases come from the same > issue in create_ordered_paths. The ordered_rel has the same path in its > pathlist > as the input_rel. Sometimes, this path is removed and freed from ordered_rel, > which leads to a dangling pointer in the child RelOptInfo. > > I've attached a patch that shows how to fix the issue. Some regression tests > change because of a hidden rule where a projection and its subpath have > different target lists. Right now, the patch always enforces a projection, > even > if the target lists are the same. This is still open for discussion on whether > there's a better way to handle it.
IMO, we should write a function like copy_path() or reparent_path(), which creates a copy of the given Path, or the latter also would copy then set the ->parent to the given RelOptInfo. Any time we use a path directly from the pathlist of another RelOptInfo, we should reparent or copy it. We could add an Assert in add_path() to check the new path has the correct parent to help us find the places where we forget to do this. David
