On Wed, Jan 11, 2017 at 1:24 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>> Well, it's not *that* consistent. If we were estimating all the numbers
>> underneath the Gather as being per-worker numbers, that would make some
>> amount of sense. But neither the other seqscan, nor the hash on it, nor
>> the hashjoin's output count are scaled that way. It's very hard to call
>> the above display anything but flat-out broken.
>
> While investigating why Rushabh Lathia's Gather Merge patch sometimes
> fails to pick a Gather Merge plan even when it really ought to do so,
> I ran smack into this problem. I discovered that this is more than a
> cosmetic issue. The costing itself is actually badly broken.
>
> The reason why this is happening is that final_cost_nestloop(),
> final_cost_hashjoin(), and final_cost_mergejoin() don't care a whit
> about whether the path they are generating is partial. They apply the
> row estimate for the joinrel itself to every such path generated for
> the join, except for parameterized paths which are a special case. I
> think this generally has the effect of discouraging parallel joins,
> because the inflated row count also inflates the join cost. I think
> the right thing to do is probably to scale the row count estimate for
> the joinrel's partial paths by the leader_contribution value computed
> in cost_seqscan.
>
> Despite my general hatred of back-patching things that cause plan
> changes, I'm inclined to think the fix for this should be back-patched
> to 9.6, because this is really a brown-paper-bag bug. If the
> consensus is otherwise I will of course defer to that consensus.
And here is a patch which seems to fix the problem.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers