Tom Lane wrote: > > I've developed the attached patch against HEAD, and no longer see any > funny behavior. Would appreciate it if you'd test some more, though. >
The patch works for me. With the patch applied, I don't see the weird errors in the pgbench and other customized tests that I used to see earlier. I looked at the patch as well. ISTM that we are now moving chains in pieces where each piece is terminated by a DEAD tuple. That implies that the MOVED_OFF chain is actually broken. This should not be a problem as long as our assumption that all RECENTLY_DEAD tuples preceding a DEAD tuple must also be DEAD and its only the way OldtestXmin is calculated that we see them as RECENTLY_DEAD. If that assumption is true (and it must be true for us to move the chain in pieces), doesn't that mean we don't really need to move the RECENTLY_DEAD tuples preceding a DEAD tuple ? One way to do so would be to collect the target tuple (the tuple from where we started following the t_ctid chain) in the free_offsets if a DEAD or INSERT_IN_PROGRESS tuple is found while following the t_ctid chain. One-by-one we would collect all the RECENTLY_DEAD tuples preceding a DEAD tuple in the truncated pages. Not that I am suggesting we do this, just wanted to check if there is a flaw in my thinking. I agree that we should not be spending too much time on fixing this corner case and the patch that you have developed is good enough. Thanks, Pavan -- EnterpriseDB http://www.enterprisedb.com ---------------------------(end of broadcast)--------------------------- TIP 2: Don't 'kill -9' the postmaster
