> > What is actually estimated wrong here seems to be the estimated
> > effective cache size, and thus the cache ratio of page fetches.
>
> Good point, but I think the estimates are only marginally sensitive
> to estimated cache size (if they're not, we have a problem, considering
> how poorly we can estimate the kernel's disk buffer size). It would
> be interesting for Daniel to try a few different settings of
> effective_cache_size and see how much the EXPLAIN costs change.
Well, the number I told him (29370) should clearly prefer the index.
The estimate is very sensitive to this value :-(
With 29370 (=229 Mb) the index cost is 1,364 instead of 3,887 with the
default of 1000 pages ==> index scan.
229 Mb file cache with 512Mb Ram is a reasonable value, I have
a lot more here:
Memory Real Virtual
free 0 MB 218 MB
procs 95 MB 293 MB
files 159 MB
total 256 MB 512 MB
Andreas