On Mon, Mar 25, 2013 at 5:57 PM, Greg Stark <stark@mit.edu> wrote:
> On Mon, Mar 25, 2013 at 9:53 PM, Kevin Grittner <kgrittn@ymail.com> wrote:
>> That would make it harder to construct a degenerate case
>
> I don't think it's hard at all. It's the same as the case Simon wants
> to solve except that the cost is incurred in a different way. Imagine
> a system where there's a huge data load to a table which is then
> read-only for an OLTP system. Until vacuum comes along -- and it may
> never since the table never sees deletes or updates -- every
> transaction needs to do a clog lookup for every tuple it sees. That
> means a significant cpu slowdown for every row lookup forever more. To
> save a one-time i/o cost.
That is simply not the case, unless every tuple was created by a
unique (or at least non-sequential) transaction xid. There is a 'last
transaction id' guard over clog lookup which is pretty effective. The
real cost is actually i/o from writing hint bits.
merlin