Re: [HACKERS] Proposal: Local indexes for partitioned table - Mailing list pgsql-hackers

From Alvaro Herrera
Subject Re: [HACKERS] Proposal: Local indexes for partitioned table
Date
Msg-id 20171215221243.3jzenwvai2dyybvn@alvherre.pgsql
Whole thread Raw
In response to Re: [HACKERS] Proposal: Local indexes for partitioned table  (Robert Haas <robertmhaas@gmail.com>)
List pgsql-hackers
Robert Haas wrote:
> On Thu, Nov 30, 2017 at 7:02 AM, Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:
> > Great question.  So you're thinking that the planner might have an
> > interest in knowing what indexes are defined at the parent table level
> > for planning purposes; but for that to actually have any effect we would
> > need to change the planner and executor also.  And one more point, also
> > related to something you said before: we currently (I mean after my
> > patch) don't mark partitioned-table-level indexes as valid or not valid
> > depending on whether all its children exist, so trying to use that in
> > the planner without having a flag could cause invalid plans to be
> > generated (i.e. ones that would cause nonexistent indexes to be
> > referenced).
> 
> Did you do it this way due to locking concerns?

No -- just because since the index-on-parent is a different kind of
object (RELKIND_PARTITIONED_INDEX) it is not considered for anything in
the planner anyway, so there's no need for indisvalid to be "correct".
Changing the flag in the parent index only needs to examine state on the
children, not modify them, so I don't think there would be any serious
locking problem.


By the way, my implementation of ALTER INDEX ATTACH PARTITION was prone
to deadlocks because it needs locks on parent table, index-on-parent,
partition, and index-on-partition; and there was no consideration to the
ordering in which these were being acquired.  I wrote a callback for
RangeVarGetRelidExtended to be called in ATExecAttachPartitionIdx() that
tries to acquire locks in the right order -- but I recently realized
that even that is not sufficient, because (unless I misread it) ALTER
INDEX itself does not worry about locking the containing table before
the index.  I think some tweaking needs to be done in
RangeVarCallbackForAlterRelation to lock the table if an index is being
altered (conditionally, depending on the precise ALTER TABLE operation).
It surprises me that we don't need to do that yet, but I haven't looked
into it any further.

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services


pgsql-hackers by date:

Previous
From: Robert Haas
Date:
Subject: Re: [HACKERS] Proposal: Local indexes for partitioned table
Next
From: Alvaro Herrera
Date:
Subject: Re: [HACKERS] Proposal: Local indexes for partitioned table