On Tue, Nov 22, 2016 at 10:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> [ Let's invent Oracle-style UNDO logs ]
>
> I dunno. I remember being told years ago, by an ex-Oracle engineer,
> that he thought our approach was better. I don't recall all the details
> of the conversation but I think his key point was basically this:
>
>> - Reading a page that has been recently modified gets significantly
>> more expensive; it is necessary to read the associated UNDO entries
>> and do a bunch of calculation that is significantly more complex than
>> what is required today.
>
> Oracle spends a lot of time on this, and it's really cache-inefficient
> because the data is spread all over. This was what this guy felt in
> circa 2001; I'd have to think that the cache unfriendliness problem is
> much worse for modern hardware.
Yes, that's the main thing I'm worried about with this design. Now,
bloat is pretty cache-inefficient too. If our table is 2x the size of
Oracle's table, they can spend more cycles per page and still do
fairly well. If that 2x increment pushes us out of RAM, they can
spend basically all the CPU time they want and still win handily. But
that isn't to say that this concern isn't justified.
> Which is not to say that it's not worth experimenting with. But what you
> describe is going to be a huge amount of work even to get to the point
> where we could try to measure the actual costs and benefits :-(
Yeah.
> Heikki's been fooling with some ideas that I think have more promise.
> I wish he'd get to the point of presenting them publicly rather than
> just over beers at conferences.
That would be good, too!
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company