Re: Freeze avoidance of very large table.

Robert Haas <robertmhaas@gmail.com> writes:
> The patch makes some attempt to update the comment mechanically, but
> that's not nearly enough.  That comment is explaining that you *can't*
> rely on the visibility map to tell you *for sure* that a page does not
> require vacuuming.  For current uses, that's OK, because if we miss a
> page we'll pick it up later.  But now we want to skip vacuuming pages
> for relfrozenxid/relminmxid advancement, that rationale doesn't apply.
> Missing pages that need to be frozen and advancing relfrozenxid anyway
> would be _bad_.

Check.

> However, after some further thought, I think we might actually be OK.
> If a page goes from all-frozen to not-all-frozen while VACUUM is
> running, any new XID added to the page must be newer than the
> oldestXmin value computed by vacuum_set_xid_limits(), so it won't
> affect the value to which we can safely set relfrozenxid.  Similarly,
> any MXID added to the page will be newer than GetOldestMultiXactId(),
> so setting relminmxid is still safe for similar reasons.

Yeah, I agree with this, as long as the issue is only that the visibility
map result is slightly stale and not that it's, say, not crash-safe.
We can reasonably assume that any newly-added XID must be one that was
in progress while VACUUM was running, and hence will be after the xmin
horizon we computed earlier.  This requires the existence of a read
barrier somewhere between computing xmin horizon and inspecting the
visibility map, but I find it hard to believe there aren't plenty.
        regards, tom lane