Tom Lane wrote: > Here's a sketch of the idea as it's developing in my mind: > > 1. We consider that relminxid = FrozenXid means that the table is > frozen, ie is guaranteed to contain no valid XIDs except FrozenXid. > Otherwise, relminxid must be a lower bound on the non-frozen XIDs in > the table. > > 2. VACUUM FREEZE acquires ExclusiveLock, vacuums the table replacing > all XIDs with FrozenXid, and if successful sets relminxid = FrozenXid. > (It might not be successful, eg it might see recently-dead tuples it > can't remove; we can't replace their xmin/xmax obviously.) In all other > cases, VACUUM sets relminxid = Min(oldest unfrozen xid in table, > transaction xmin) (or some other convenient lower-bound computation, > eg maybe just use the cutoff instead of actively figuring the min XID). > I think we have to XLOG the setting of relminxid to be safe. > > 3. Any modification of a table (that inserts an XID into it) must check > to see if relminxid = FrozenXid, and if so change it to transaction xmin > (or some other lower bound on the oldest running XID). This action has > to be WAL-logged. > > 4. VACUUM has to recompute datminxid to be the oldest non-frozen > relminxid in the database (but not more than transaction xmin, to cover > case where someone else is creating a table concurrently). We might be > able to go back to your idea of not having to do this work unless the > prior value of relminxid matches datminxid. I think plain VACUUM could > skip tables having relminxid = FrozenXid altogether. > > I'm tempted to say that the "unfreezing" action (#3) could just be done > at the point where we open a rel and take a stronger-than-AccessShare > lock on it. This would minimize the overhead needed, and give us pretty > good confidence we'd not missed any places. It'd mean that, say, an > UPDATE that changed no rows would still mark the table unfrozen, but I > see no great downside to that.
Seems fine. I'll devote some time to this. -- Alvaro Herrera http://www.CommandPrompt.com/ PostgreSQL Replication, Consulting, Custom Development, 24x7 support ---------------------------(end of broadcast)--------------------------- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match