Re: Autovacuum improvements

Alvaro Herrera <alvherre(at)commandprompt(dot)com> writes:
> I'm cooking a patch for this which seems pretty reasonable, but I'm
> having a problem: what mechanism do we have for waiting until a process
> exits?

None, and I think you probably don't want to sit on the database lock
while waiting, either. I was envisioning a simple sleep loop, viz

for(;;)
{
acquire database lock;
foreach(PGPROC entry in that database)
{
if (it's autovac)
send sigint;
else
fail;
}
if (found any autovacs)
{
release database lock;
sleep(100ms or so);
/* loop back and try again */
}
else
break;
}

Also see Peter's nearby suggestion that we ought to wait instead of fail
for *all* cases of somebody attached to the database. This would adapt
readily enough to that.

I was complaining elsewhere that I didn't want to use a sleep loop
for fixing the fsync-synchronization issue, but CREATE/DROP DATABASE
seems a much heavier-weight operation, so I don't feel that a sleep
is inappropriate here.

> Maybe make autovacuum acquire an LWLock at start, which it then
> keeps until it's gone, but it seems wasteful to have a lwlock just for
> that purpose.

And it doesn't scale to multiple autovacs anyway, much less the wait-for-
everybody variant.

regards, tom lane