On Thu, Mar 17, 2016 at 2:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Well, I'm prepared to yield to the extent of repeating the hook call
> before each phase with an UpperRelationKind parameter to tell which phase
> we're about to do. The main concern here is to avoid redundant
> computation, but the hook can check the "kind" parameter and fall out
> quickly if it has nothing useful to do at the current phase.
>
> I do not, however, like the proposal to expose wflists and so forth.
> Those are internal data structures in grouping_planner and I absolutely
> refuse to promise that they're going to stay stable. (I had already
> been thinking a couple of weeks ago about revising the activeWindows
> data structure, now that it would be reasonably practical to cost out
> different orders for doing the window functions in.) I think a hook
> that has its own ideas about window function implementation methods
> can gather its own information about the WFs without that much extra
> cost, and it very probably wouldn't want exactly the same data that
> create_window_paths uses today anyway.
>
> So what I would now propose is
>
> typedef void (*create_upper_paths_hook_type) (PlannerInfo *root,
> UpperRelationKind stage,
> RelOptInfo *input_rel);
>
> and have this invoked at each stage right before we call
> create_grouping_paths, create_window_paths, etc.
Works for me.
> Also, I don't particularly see a need for a corresponding API for FDWs.
> If an FDW is going to do anything in this space, it presumably has to
> build up ForeignPaths for all the steps anyway. So I'd be inclined
> to leave GetForeignUpperPaths as-is.
No idea if that is going to be a significant limitation or not.
Doesn't seem like it should be, but what do I know?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company