Re: Performance regression between 8.3 and 8.4 on heavy text indexing

2009/8/27 Gaël Le Mignot <gael@pilotsystems.net>:
> The  weird thing  was  that with  the  default of  100 for  statistics
> target, it was  worse than when we  moved back to 10. So  I didn't try
> with 1000, but I should have.

When you have so much data and a statistics target so low, you can't
expect the sample taken to be representative :between different runs
of ANALYZE, you can have totally different estimates and so totally
different plans. You just were lucky at 10 and unlucky at 100.

Thanks for your feedback and it's nice to see your problem solved.

--
Guillaume