Bruce Momjian wrote:
> > A different approach that's been discussed on pghackers is to make use
> > of btree indexes for columns that have such indexes: we could scan the
> > indexes to visit all the column values in sorted order. I have rejected
> > that approach because (a) it doesn't help for columns without a suitable
> > index; (b) our indexes don't distinguish deleted and live tuples,
> > which would skew the statistics --- in particular, we couldn't tell a
> > frequently-updated single tuple from a commonly repeated value; (c)
> > scanning multiple indexes would likely require more total I/O than just
> > grabbing sample tuples from the main table --- especially if we have to
> > do that anyway to handle columns without indexes.
>
> Remember one idea is for index scans to automatically update the expired
> flag in the index bitfields when they check the heap tuple.
And we should really do that. While playing around with my (for 7.2 to be) access statistics stuff I found
that when running pg_bench, a couple of thousand index scans cause millions and millions of buffer
fetches, because that pg_bench updates one and the same row over and over again and it has a PKEY.
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck@Yahoo.com #
_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com
