On Tue, Mar 12, 2024 at 3:36 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Yeah. I keep thinking that the right solution is to not generate
> disabled paths in the first place if there are any other ways to
> produce the same relation. That has obvious order-of-operations
> problems though, and I've not been able to make it work.
I've expressed the same view in the past. It would be nice not to
waste planner effort on paths that we're just going to throw away, but
I'm not entirely sure what you mean by "obvious order-of-operations
problems."
To me, it seems like what we'd need is to be able to restart the whole
planner process if we run out of steam before we get done. For
example, suppose we're planning a 2-way join where index and
index-only scans are disabled, sorts are disabled, and nested loops
and hash joins are disabled. There's no problem generating just the
non-disabled scan types at the baserel level, but when we reach the
join, we're going to find that the only non-disabled join type is a
merge join, and we're also going to find that we have no paths that
provide pre-sorted input, so we need to sort, which we're also not
allowed to do. If we could give up at that point and restart planning,
disabling all of the plan-choice constraints and now creating all
paths for each RelOptInfo, then everything would, I believe, be just
fine. We'd end up needing neither disable_cost nor the mechanism
proposed by this patch.
But in the absence of that, we need some way to privilege the
non-disabled paths over the disabled ones -- and I'd prefer to have
something more principled than disable_cost, if we can work out the
details.
--
Robert Haas
EDB: http://www.enterprisedb.com