Re: performance issue in remove_from_unowned_list()

On 3/10/19 9:09 PM, Alvaro Herrera wrote:
> On 2019-Feb-07, Tomas Vondra wrote:
>
>> Attached is a WIP patch removing the optimization from DropRelationFiles
>> and adding it to smgrDoPendingDeletes. This resolves the issue, at least
>> in the cases I've been able to reproduce. But maybe we should deal with
>> this issue earlier by ensuring the two lists are ordered the same way
>> from the very beginning, somehow.
>
> I noticed that this patch isn't in the commitfest. Are you planning to
> push it soon?
>

I wasn't planning to push anything particularly soon, for two reasons:
Firstly, the issue is not particularly pressing except with non-trivial
number of relations (and I only noticed that during benchmarking).
Secondly, I still have a feeling I'm missing something about b4166911
because for me that commit does not actually fix the issue - i.e. I can
create a lot of relations in a transaction, abort it, and observe that
the replica actually accesses the relations in exactly the wrong order.
So that commit does not seem to actually fix anything.

Attached is a patch adopting the dlist approach - it seems to be working
quite fine, and is a bit cleaner than just slapping another pointer into
the SMgrRelationData struct. So I'd say this is the way to go.

I see b4166911 was actually backpatched to all supported versions, on
the basis that it fixes oversight in 279628a0a7. So I suppose this fix
should also be backpatched.

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services