Re: pg_upgrade < 9.3 -> >=9.3 misses a step around multixacts

On 2014-07-20 15:55:25 -0400, Tom Lane wrote:
> I wrote:
> > I started transcribing Bruce's proposed fix procedure at
> > https://wiki.postgresql.org/wiki/20140702pg_upgrade_fix
> > into the release notes, but I'm afraid it's all wet.
>
> > He's suggesting copying the last checkpoint's NextMultiXactId into
> > datminmxid/relminmxid, which is surely the wrong thing: that's likely to
> > be newer than all mxids in the tables, not older than them.  I thought at
> > first that this was a simple thinko and he meant to write oldestMultiXid,
> > but here's the thing: if we're in the situation where we've got
> > wraparound, isn't oldestMultiXid going to be 1?  The value recorded in the
> > checkpoint isn't magic, it's just going to be extracted from whatever's in
> > pg_database; and the whole problem here is that we can't trust that data.
> > Where can we get a useful lower bound from?
>
> Ugh: it's worse than that.  pg_upgrade itself is using this utterly
> nonsensical logic to set datminmxid/relminmxid.  This is a stop-ship
> issue for 9.3.5.
>
> After some reflection it seems to me that we could estimate oldestmxid for
> a pre-9.3 source cluster as the NextMultiXactId from its pg_control less
> 2000000000 or so.  This will nearly always be much older than the actual
> oldest mxid, but that's okay --- the next vacuuming cycle will advance the
> datminmxid/relminmxid values to match reality, so long as they aren't
> wrapped around already.

I think NextMultiXactId should be fine when upgrading from pre 9.3
clusters. Pre 9.3 multis that have been written before a shutdown don't
really have a meaning: To the point that we actually never look at
them. MultiXactIdSetOldestVisible() would have set oldestMXact to
MultiXactState->nextMXact if no backend has any active
ones. GetMultiXactIdMembers et al will deal with that.

So I think for pg_upgrade itself using NextMulti as a replacement for
oldestMXact is fine.

Now that obviously does *not* mean it's safe to do that to after the
cluster has already been running 9.3+.

Greetings,

Andres Freund

--
 Andres Freund                       http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services