Обсуждение: Introducing the future Debian multiversion/multicluster architecture

Hi!

As an ultra-short introduction, I am the primary Debian developer for
PostgreSQL packages.

Our current packages became pretty hard to maintain; in particular
automatic upgrades to new major versions have always been a pain and
fail very often. There are also many requests for supporting more than
one cluster.

Thus we are currently developing a completely new packaging system and
architecture for future PostgreSQL packages which allow to manage
arbitrarily many clusters (also of different versions) in parallel.
This allows to do cluster updates in a running system, independently
of package upgrades, and will generally be much more robust and useful
also for web hosters.

I have heard that some guys of other distributions want to do
something similar in the future, so I was asked to publish the general
ideas for public review and discussion:

  http://people.debian.org/~mpitt/postgresql-ng.html

Thanks in advance for any thoughts,

Martin

(Please respect Mail-Followup-To or keep me CC'ed, I'm not
subscribed).

--
Martin Pitt                       http://www.piware.de
Ubuntu Developer            http://www.ubuntulinux.org
Debian GNU/Linux Developer       http://www.debian.org

Martin,

This looks really good.  I wish it were going into Sarge, though of
course the timing isn't right for that.  :)

A couple things I noticed about the automated upgrade procedure
(pg_version_upgrade):

1) Since it uses pg_dumpall, it doesn't seem to be capable of handling
   databases with large objects (these have to be dumped individually
   using one of the non-text dump formats and the -b option).
   Furthermore, I believe it will fail silently, as pg_dumpall will
   simply leave out the pg_largeobject table and return success for
   such databases, which will be left with dangling LO references.

2) It is fairly commonly recommended, when doing a dump/restore
   upgrade, to use the new version's pg_dump to dump out the old
   database, as that generally makes it easier to load into the new
   version without manual surgery on the dump file.  The sentence "All
   operations are done with the software version appropriate to the
   cluster version." in the document seems to preclude this option.

Thanks for putting so much thought into this!

-Doug