On Fri, 14 Feb 2003, Bruce Momjian wrote:
> OK, once we have PITR, will anyone want incremental backups?
Well, I'm not entirely clear on how PITR will work, so I may be off-base
here, but it seems to me that offering incremental backups that back
up only changed pages might not be all that big a win, given how
postgres writes its pages. On DBMSs that don't use MVCC, if you change a
particular item in a row ten times, one page is changed. If you do it in
postgres, you could well be changing ten pages, as the system writes the
two copies of the entire row wherever it can find space. So in databases
where a lot of rows are changed, where an incremental backup would
normally be a win because it would be much smaller than the logs over a
given period, it isn't going to be with postgres.
But you know, if we could get rid of redundant changes in the logs we're
using for backup, that could save a lot of space in a situation like
the one I described above. If a particular row and column is changed
fifty times over the course of a month, it's going to be recorded fifty
times in the log. But there's really no need for all fifty of those,
if you don't mind not being able to restore to any time before the
current time. You can reduce the size of the logs you need to store
for backup by throwing away the first forty-nine of those changes, and
keeping only the most recent version. There shouldn't be any worries
about referential integrity, because when you do a restore, you start
with a full backup that is ok, and once you've successfully applied
all the transactions in the log, you know it will be ok again, so any
intermediate states during the restore where integrity is not maintained
are not a problem.
cjs
--
Curt Sampson <cjs@cynic.net> +81 90 7737 2974 http://www.netbsd.org Don't you know, in this new Dark Age, we're
alllight. --XTC