Re: tuning autovacuum

On 06/09/2011 05:41 PM, Tom Lane wrote:
> As Robert said, we're already seeing scalability problems with the
> pg_stats subsystem.  I'm not eager to add a bunch more per-table
> counters, at least not without some prior work to damp down the ensuing
> performance hit.
>    

That's fair.  Anyone who is running into the sort of autovacuum issues 
prompting this discussion would happily pay the overhead to get better 
management of that; it's one of the easiest things to justify more 
per-table stats on IMHO.  Surely the per-tuple counters are vastly more 
of a problem than these messages could ever be.

But concerns about stats overload are why I was highlighting issues 
around sending multiple messages per vacuum, and why incremental updates 
as it runs are unlikely to work out.  Balancing that trade-off, getting 
enough data to help but not so such the overhead is obnoxious, is the 
non obvious tricky part of the design here.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us