On Tue, 2007-11-13 at 14:36 -0500, Greg Smith wrote:
> On Tue, 13 Nov 2007, Andrew Sullivan wrote:
>
> > I have to agree with what Tom says, however, about people thinking
> > they're smarter than the system. Much of the time, this sort of thumb
> > on the scale optimisation just moves the cost to some other place
>
> Sure, but in this case the reasoning seems sound enough. The buffer
> eviction policy presumes that all buffers cost an equal amount to read
> back in again. Here we have an application where it's believed that's not
> true: the data on disk for this particular table has a large seek
> component to it for some reason, it tends to get read in large chunks (but
> not necessairly frequently), and latency on that read is critical to
> business requirements. "The system" doesn't know that, and it's
> impractical to make it smart enough to figure it out on its own, so asking
> how to force that is reasonable.
It seems possible to imagine a different buffer eviction policy based
upon tablespace, block type, peak rather than latest usage pattern etc..
-- Simon Riggs 2ndQuadrant http://www.2ndQuadrant.com