On Thu, 2007-10-04 at 17:33 -0400, Alvaro Herrera wrote: > Simon Riggs escribió: > > > Seems like we don't need to mess with the deadlock checker itself. > > > > We can rely on the process at the head of the lock wait queue to sort > > this out for us. So all we need do is look at the isAutovacuum flag on > > the process that is holding the lock we're waiting on. If it isn't an > > autoANALYZE we can carry on with the main deadlock check. We just need a > > new kind of deadlock state to handle this, then let ProcSleep send > > SIGINT to the autoANALYZE and then go back to sleep, waiting to be > > reawoken when the auotANALYZE aborts. > > Ok, I think this makes sense. > > I can offer the following patch -- it makes it possible to determine > whether an autovacuum process is doing analyze or not, by comparing the > PGPROC of the running WorkerInfo list (the list has at most > max_autovacuum_workers entries, so this is better than trolling > ProcGlobal).
OK, I've got this working now. It successfully handles this test case, which trips up on an auto ANALYZE every time I run it. -- drop table a; create table a as select generate_series(1,1000000)::integer as col1; alter table a alter column col1 type bigint; alter table a alter column col1 type bigint; alter table a alter column col1 type bigint; alter table a alter column col1 type bigint; alter table a alter column col1 type bigint; -- I think there may be a cleaner implementation, so I'll clean it up and post tomorrow. Few thoughts: Why do we run all of the ANALYZEs in a single big transaction? That seems like it could be the cause of many problems. ANALYZE specifically holds locks until EOXact, so I'd recommend we start a new transaction for each one. What do you think? I notice when we cancel an AV worker it always says "cancelling autovacuum of table", even when its just an ANALYZE. Wasn't important before but now looks a little strange. If you want to commit this patch, I'll layer mine over the top. Any other input anyone? -- Simon Riggs 2ndQuadrant http://www.2ndQuadrant.com ---------------------------(end of broadcast)--------------------------- TIP 4: Have you searched our list archives? http://archives.postgresql.org