On Wed, Sep 14, 2005 at 11:30:52PM -0400, Tom Lane wrote:
> Updating the pg_database entry
> will have to be the responsibility of VACUUM itself. It's not that
> terrible: you don't have to scan pg_class unless you see that the
> pg_class.relfreezexid value you are replacing is equal to
> pg_database.datfreezexid, and with the exact computation suggested
> above, that won't be a common occurrence.
Ok, this patch is a first cut at that. I made it slightly smarter,
because we want to do it only once if we issue a database-wide vacuum,
instead of once per table. However, I also had to cope with the
possibility that the table with the minimum Xid value is dropped, so I
made that set the datminxid to InvalidXid. If after a VACUUM the
datminxid is found to be Invalid, pg_class is scanned inconditionally.
There's the usual gotcha with shared relations: vacuuming them on one
database is not going to update their pg_class entries in other
databases, so they will be vacuumed more frequently than really
necessary. I don't see how to fix that -- one way would be storing
their stats in one database only, but it seems dangerous.
Note that I use LockSharedObject() to lock the database while we are
updating the pg_database row. This means that more than one database
can be updated concurrently (this is important because we have to keep
the lock while we seqscan pg_class). This may be a bad idea from the
point of view of the buffer manager; maybe we need an additional
LockBuffer() just before we are going to modify the tuple.
With this in place, it's no longer necessary to issue database-wide
vacuums anymore. Note that I haven't tested the part where the clog
grows enough to be truncated, nor really anything more complicated than
a single backend doing database-wide or single-table vacuums (with an
optional parallel backend with an open transaction). Still, it shows
what the basics of the patch are.
--
Alvaro Herrera http://www.amazon.com/gp/registry/CTMLCN8V17R4
"Lo esencial es invisible para los ojos" (A. de Saint Exúpery)
