Joseph S wrote:
> Greg Smith wrote:
>> Joseph S wrote:
>>> So I run "select count(*) from large_table" and I see in xosview a
>>> solid block of write activity. Runtime is 28125.644 ms for the first
>>> run. The second run does not show a block of write activity and
>>> takes 3327.441 ms
>> http://wiki.postgresql.org/wiki/Hint_Bits
>>
>
> Hmm. A large select results in a lot of writes? This seems broken.
> And if we are setting these hint bits then what do we need VACUUM
> for? Is there any way to tune this behavior? Is there any way to get
> stats on how many rows/pages would need hint bits set?
Basically, the idea is that if you're pulling a page in for something
else that requires you to compute the hint bits, just do it now so
VACUUM doesn't have to later, while you're in there anyway. Why make
VACUUM do the work later if you're doing part of it now anyway? If you
reorganize your test to VACUUM first *before* running the "select (*)
from...", you'll discover the writes during SELECT go away. You're
running into the worst-case behavior. For example, if you inserted a
bunch of things more slowly, you might discover that autovacuum would do
this cleanup before you even got to looking at the data.
There's no tuning for the behavior beyond making autovacuum more
aggressive (to improve odds it will get there first), and no visibility
into what's happening either. And cranking up autovacuum has its own
downsides. This situation shows up a lot when you're benchmarking
things, but not as much in the real world, so it's hard to justify
improving.
--
Greg Smith 2ndQuadrant Baltimore, MD
PostgreSQL Training, Services and Support
greg@2ndQuadrant.com www.2ndQuadrant.com