Date: Wed, 7 Mar 2018 16:48:11 -0300
From: Alvaro Herrera <alvherre@2ndquadrant.com>
To: bricklen <bricklen@gmail.com>
Cc: Natalie Wenz <nataliewenz@ebureau.com>,
	pgsql-admin <pgsql-admin@postgresql.org>
Subject: Re: pg_upgrade and frozen xids
Message-ID: <20180307194811.tloezvberxo33faw@alvherre.pgsql>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: 
 <CAGrpgQ_pUDdNXTFZotYoOR-iW-OXtTbFZ-69u-1=4dM+kWVPnA@mail.gmail.com>
User-Agent: NeoMutt/20170306-137-4415bd-dirty (1.8.0)
Precedence: bulk

bricklen wrote:

> We recently upgraded about 1500 Postgres clusters from 9.3 to 10.1 and we
> are seeing similar issues. Specifically:
> * An autovacuum worker stuck vacuuming regular tables for days until they
> are killed (I've seen "vacuuming indexes" phase), and they are showing 99%
> CPU in top. pg_cancel|terminate_backend does not usually kill the process,
> only a hard kill does. In the latter case, two (out of four) times it also
> crashed Postgres.
> * There have been about 10 examples of btree index corruption, across a
> handful of databases, in different data centres. In two of the cases, it's
> possible there were hardware issues as the hypervisor those two VM's were
> running on crashed.

I wonder if this may be related to changes in collation algorithm.  If
the indexes are all on text columns, I think you would be better served
by reindexing them all instead of trying to treat this situation as a
code bug.

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