Bruce Momjian wrote:
>
> mlw wrote:
> > Now, given the choice of the two strategies on a table, both pretty close to
> > one another, the risk of poor performance for using the index scan is minimal
> > based on the statistics, but the risk of poor performance for using the
> > sequential scan is quite high on a large table.
> My second point, that index scan is more risky than sequential scan, is
> outlined above. A sequential scan reads each page once, and uses the
> file system read-ahead code to prefetch the disk buffers. Index scans
> are random, and could easily re-read disk pages to plow through a
> significant portion of the table, and because the reads are random,
> the file system will not prefetch the rows so the index scan will have
> to wait for each non-cache-resident row to come in from disk.
That is a very interesting point, but shouldn't that be factored into the cost
(random_tuple_cost?) In which case my point still stands.
