Re: Pg_upgrade speed for many tables

On Mon, Nov  5, 2012 at 06:33:16PM -0300, Alvaro Herrera wrote:
> Bruce Momjian escribió:
> > On Mon, Nov  5, 2012 at 04:14:47PM -0500, Robert Haas wrote:
> > > On Mon, Nov 5, 2012 at 4:07 PM, Jeff Janes <jeff.janes@gmail.com> wrote:
> > > > Or have options for pg_dump and pg_restore to insert "set
> > > > synchronous_commit=off" into the SQL stream?
> > > 
> > > It would be kind of neat if we had a command that would force all
> > > previously-asynchronous commits to complete.  It seems likely that
> > > very, very few people would care about intermediate pg_dump states, so
> > > we could do the whole dump asynchronously and then do "FORCE ALL
> > > COMMITS;" or whatever at the end.
> > 
> > Actually, I had assumed that a session disconnection forced a WAL fsync
> > flush, but now I doubt that.  Seems only server shutdown does that, or a
> > checkpoint.  Would this work?
> > 
> >     SET synchronous_commit=on;
> >     CREATE TABLE dummy(x int);
> >     DROP TABLE dummy;
> 
> AFAIR any transaction that modifies catalogs gets sync commit forcibly,
> regardless of the setting.  And sync commit means you get to wait for

Uh, I am not seeing that my testing because I was only doing CREATE
TABLE and it was affected by the synchronous_commit value.

> all previous transactions to be flushed as well.  So simply creating a
> temp table ought to do the trick ...

I don't think TEMP tables write to WAL, for performance reasons.

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