Re: First steps with 8.3 and autovacuum launcher

Simon Riggs schreef:
> On Fri, 2007-10-12 at 11:44 +0200, Michael Paesold wrote:
>   
>> Simon Riggs wrote:
>>     
>>> On Fri, 2007-10-12 at 01:24 -0400, Alvaro Herrera wrote:
>>>       
>>>> Yes, I think it is easy to mark the "is for xid wraparound" bit in the
>>>> WorkerInfo struct and have the cancel work only if it's off.
>>>>
>>>> However, what I think should happen is that the signal handler for
>>>> SIGINT in a worker for xid wraparound should not cancel the current
>>>> vacuum.  Instead turn it into a no-op, if possible.  That way we also
>>>> disallow a user from cancelling vacuums for xid wraparound.  I think he
>>>> can do that with pg_cancel_backend, and it could be dangerous.
>>>>         
>>> I think that is dangerous too because the user may have specifically
>>> turned AV off. That anti-wraparound vacuum might spring up right in a
>>> busy period and start working its way through many tables, all of which
>>> cause massive writes to occur. That's about as close to us causing an
>>> outage as I ever want to see. We need a way through that to allow the
>>> user to realise his predicament and find a good time to VACUUM. I never
>>> want to say to anybody "nothing you can do, just sit and watch, your
>>> production system will be working again in no time. Restart? no that
>>> won't work either."
>>>       
>> You are probably right that VACUUM going full-steam is a bad idea in most 
>> situations. Except for anti-wraparound vacuum, cancellation seems the most 
>> reasonable thing to do. Because autovacuum will usually pickup the table in 
>> time again.
>>     
>
> Yeh, if we do have to do the second emergency anti-wraparound, then that
> should be at full speed, since there's nothing else to do at that point.
>
>   
>> The only problem I would see is if someone has an application that does a 
>> lot of schema changes (doesn't sound like a good idea anyway). In that case 
>> they would better issue manual vacuums on such tables.
>>     
>
> I can't see a use case for regular DDL as part of an application, on an
> otherwise integral table (lots of updates and deletes).
>   
As part of an application there's no use.
As part of an upgrade between 2 different versions of that application 
there is.
And that's exactly the kind of situation where temporary disabling 
autovacuum could become handy.