On Mon, Nov 5, 2012 at 03:08:17PM -0500, Bruce Momjian wrote:
> Magnus reported that a customer with a million tables was finding
> pg_upgrade slow. I had never considered many table to be a problem, but
> decided to test it. I created a database with 2k tables like this:
>
> CREATE TABLE test1990 (x SERIAL);
>
> Running the git version of pg_upgrade on that took 203 seconds. Using
> synchronous_commit=off dropped the time to 78 seconds. This was tested
> on magnetic disks with a write-through cache. (No change on an SSD with
> a super-capacitor.)
>
> I don't see anything unsafe about having pg_upgrade use
> synchronous_commit=off. I could set it just for the pg_dump reload, but
> it seems safe to just use it always. We don't write to the old cluster,
> and if pg_upgrade fails, you have to re-initdb the new cluster anyway.
>
> Patch attached. I think it should be applied to 9.2 as well.
Modified patch attached and applied to head and 9.2. I decided to use
synchronous_commit=off only on the new cluster, just in case we ever do
make a modification of the old cluster.
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://enterprisedb.com
+ It's impossible for everything to be true. +