Alvaro,
* Alvaro Herrera (alvherre@alvh.no-ip.org) wrote:
> Greg Stark wrote:
>
> > The general shape of what I would like to see is some log which lists
> > where each checkpoint starts and ends and what blocks are modified
> > since the previous checkpoint. Then to generate an incremental backup
> > from any point in time to the current you union all the block lists
> > between them and fetch those blocks. There are other ways of using
> > this aside from incremental backups on disk too -- you could imagine a
> > replica that has fallen behind requesting the block lists and then
> > fetching just those blocks instead of needing to receive and apply all
> > the wal.
>
> Hmm, this sounds pretty clever. And we already have the blocks touched
> by each record thanks to the work for pg_rewind (so we don't have to do
> any nasty tricks like the stuff Suzuki-san did for pg_lesslog, where
> each WAL record had to be processed individually to know what blocks it
> referenced), so it shouldn't be *too* difficult ...
Yeah, it sounds interesting, but I was just chatting w/ David about it
and we were thinking about how checkpoints are really rather often done,
so you end up with quite a few of these lists being out there.
Now, if the lists were always kept in a sorted fashion, then perhaps we
would be able to essentially merge-sort them all back together and
de-dup that way, but even then, you're talking about an awful lot if
you're looking at daily incrementals- that's 288 standard 5-minute
checkpoints, each with some 128k pages changed, assuming max_wal_size of
1GB, and I think we can all agree that the default max_wal_size is for
rather small systems. That ends up being something around 2MB per
checkpoint to store the pages in or half a gig per day just to keep
track of the pages which changed in each checkpoint across that day.
There's a bit of hand-waving in there, but I don't think it's all that
much to reach a conclusion that this might not be the best approach.
David and I were kicking around the notion of a 'last LSN' which is kept
on a per-relation basis, but, of course, that ends up not really being
granular enough, and would likely be a source of contention unless we
could work out a way to make it "lazy" updated somehow, or similar.
> > It would also be useful for going in the reverse direction: look up
> > all the records (or just the last record) that modified a given block.
>
> Well, a LSN map is what I was suggesting.
Not sure I entirely followed what you were getting at here..?
Thanks!
Stephen
