Re: [HACKERS] Duplicated row after promote in synchronous streaming replication

On 26 March 2014 15:08, Dang Minh Huong <kakalot49(at)gmail(dot)com> wrote:
> Hi all,
>
> I'm using PostgreSQL 9.1.10 for my HA project and have found this problem.
>
> I did (multiple times) the following sequence in my primary/standby
> synchronous replication environment,
>
> 1. Update rows in a table (which have primary key constraint column) in
> active DB
>
> 2. Stop active DB
>
> 3. Promote standby DB
>
> 4. Confirm the updated table in promoted standby (new primary) and found
> that, there's a duplicate updated row (number of row was increased).
>
> I think it is a replication bug but wonder if it was fixed yet.
> Can somebody help me?
>
> I'm not yet confirm PostgreSQL source, but here is my investigation result.
>
> Updated table before promoted were HOT update (index file was not changed).
>
> After promote i continue update that duplicated row (it returned two row
> updated), and confirm with pg_filedump, i found the duplicated row and only
> one is related to primary key index constraint.
>
> Compare with old active DB, i saw that after promote line pointer of updated
> row (duplicated row) is broken into two line pointer, the new one is related
> to primary index constraint and the other is not related to. Some thing like
> below,
>
> Old active DB:
> ctid(0,3)->ctid(0,6)->ctid(0,7)
>
> New active DB (after promote and update):
> ctid(0,3)->ctid(0,9)
> ctid(0,7)->ctid(0,10)
>
> ctid(0,10) is not related to primary key index constraint.
>
> Is something was wrong in redo log in standby DB? Or line pointer in HOT
> update feature?

It sounds like you're hitting a bug that was introduced in that
exact minor version, and has since been fixed:

http://www.postgresql.org/docs/9.1/static/release-9-1-11.html

You should update to the latest minor version, then re-base your
standbys from the primary.

--
Thom