It seems that we're converging on the conclusion that not truncating
clog early is the least bad alternative. This has the advantage of
making things a lot simpler --- we won't need to track minxid at all.
Allow me to summarize what I think has to happen:
* VACUUM will determine a freeze cutoff XID the same way it does now,
except that instead of using a hard-wired freeze window of 1G
transactions, we'll either reduce the window to (say) 100M transactions
or provide a GUC variable that can be adjusted over some reasonable
range.
* All XIDs present in the table that are older than the cutoff XID will
be replaced by FrozenXid or InvalidXid as required, and such actions
will be WAL-logged. (I think we need to consider all 3 of xmin, xmax,
and xvac here.)
* On successful completion, the cutoff XID is stored in
pg_class.relvacuumxid, and pg_database.datvacuumxid is updated
if appropriate. (The minxid columns are now useless, but unless there
is another reason to force initdb before 8.2, I'm inclined to leave them
there and unused. We can remove 'em in 8.3.)
* pg_clog is truncated according to the oldest pg_database.datvacuumxid.
We should WAL-log this action, because replaying such an entry will
allow a PITR slave to truncate its own clog and thereby avoid wraparound
conflicts. Note that we no longer need a checkpoint before truncating
--- what we need is XLogFlush, instead. ("WAL before data")
These changes get us back into the regime where the hint bits truly are
hints, because the underlying pg_clog data is still there, both in a
master database and in a PITR slave. So we don't need to worry about
WAL-logging hint bits. We also avoid needing any flushes/fsyncs or
extra checkpoints. The added WAL volume should be pretty minimal,
because only tuples that have gone untouched for a long time incur extra
work. The added storage space for pg_clog could be annoying for a small
database, but reducing the freeze window ameliorates that objection.
Comments? Anyone see any remaining holes?
regards, tom lane
