Re: Reducing tuple overhead

Robert Haas wrote:
> On Mon, Apr 27, 2015 at 5:01 PM, Jim Nasby <Jim(dot)Nasby(at)bluetreble(dot)com> wrote:
> > The problem with just having the value is that if *anything* changes between
> > how you evaluated the value when you created the index tuple and when you
> > evaluate it a second time you'll corrupt your index. This is actually an
> > incredibly easy problem to have; witness how we allowed indexing
> > timestamptz::date until very recently. That was clearly broken, but because
> > we never attempted to re-run the index expression to do vacuuming at least
> > we never corrupted the index itself.
>
> True. But I guess what I don't understand is: how big a deal is this,
> really? The "uncorrupted" index can still return wrong answers to
> queries. The fact that you won't end up with index entries pointing
> to completely unrelated tuples is nice, but if index scans are missing
> tuples that they should see, aren't you still pretty hosed?

As I recall, there's a desire not to run expressions when vacuuming,
because to run them means getting a snapshot, in case any functions look
at database state; and you can't get a snapshot while vacuuming because
that means the vacuum gets visible to concurrent processes; in
particular they keep all processes' xmin from moving forward.

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services