Re: pg_upgrade problem with invalid indexes

On Tue, Dec 11, 2012 at 03:12:37PM -0500, Bruce Momjian wrote:
> On Fri, Dec  7, 2012 at 04:49:19PM -0500, Bruce Momjian wrote:
> > On Fri, Dec  7, 2012 at 10:38:39PM +0100, Andres Freund wrote:
> > > On 2012-12-07 16:30:36 -0500, Bruce Momjian wrote:
> > > > On Fri, Dec  7, 2012 at 04:21:48PM -0500, Tom Lane wrote:
> > > > > Andres Freund <andres@2ndquadrant.com> writes:
> > > > > > On 2012-12-07 13:59:41 -0500, Tom Lane wrote:
> > > > > >> indisvalid should be sufficient.  If you try to test more than that
> > > > > >> you're going to make the code more version-specific, without actually
> > > > > >> buying much.
> > > > >
> > > > > > Doesn't the check need to be at least indisvalid && indisready? Given
> > > > > > that 9.2 represents !indislive as indisvalid && !indisready?
> > > > >
> > > > > Um, good point.  It's annoying that we had to do it like that ...
> > > >
> > > > So, does this affect pg_upgrade?  Which PG versions?
> > > 
> > > Only 9.2 :(. Before that there was no DROP INDEX CONCURRENTLY and in 9.3
> > > there's an actual indislive field and indisready is always set to false
> > > there if indislive is false.
> > > 
> > > But I see no problem using !indisvalid || !indisready as the condition
> > > in all (supported) versions.
> > 
> > OK, updated patch attached.
> 
> Patch applied back to 9.0.
> 
> Now that it is applied, I need to publicize this.  How do I do that? 
> Josh mentioned my blog.  
> 
> What would cause these invalid indexes?  Just CREATE INDEX CONCURRENTLY
> failures?  What types of failures would users have if these invalid
> indexes had been upgraded by pg_upgrade?  Can they test their indexes in
> any way?  I assume they can't run queries on the old cluster to check.

Blog entry posted explaining the bug and fix:
http://momjian.us/main/blogs/pgblog/2012.html#December_14_2012

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +