Re: bzr commit into mysql-5.0-bugteam branch (alfranio.correia:2705) Bug#40337

[He Zhenxing]

alfranio correia wrote:
> He Zhenxing wrote:
> > Lars Thalmann wrote:
> >   
> >> On Mon, Nov 03, 2008 at 03:34:30PM +0800, He Zhenxing wrote:
> >>     
> >>> 3) Try to reduce the relay logs need to be retrieved, always fsync the
> >>> relay log file when rotating to a new file, by this way, we can be sure
> >>> that only the last relay log file can be corrupted, so there's no need
> >>> to retrieve logs before the events in the last relay log.
> >>>       
> >> Even though this is an improvement when the slave is restarted, 
> >> I think we do not need to fix this now.
> >>
> >> Motivation:
> >> - We already have the other code tested and in custom build
> >> - The recovery mechanism will only execute when there is slave crash,
> >>   so this optimization is not that useful
> >>     
> >
> > If my understanding is correct, the recovery mechanism is executed evert
> > time the slave is restarted, it does not detect for possible crashes, I
> > also mentioned this in previous comments. But I agree that if we only
> > run the recovery when we detected crashes, then the improvement to
> > reduce log events to retrieve is not that necessary.
> >   
> Let's define restart.
> 
> restart = when the mysqld is started.

Yes, this is what I mean by restart

> So if this is done manually, the user may disable the parameter.
> 

probably, but user might put this option in the configure file, or using
a script

> restart != when the user runs the command "start slave".
> So this operation does not trigger the recovery process.
> 

right

> For now, it does not detected a "normal shutdown and startup" from a "crash
> and startup". This is a new feature proposed in WL#4621
> 

I'd prefer we can add the carsh detect mechanism in this patch, but I'm
OK to ship the patch without it if implement the crash detection is not
easy or may have it's own risk. 

> Most likely I will have a draft implementation on Monday of this new
> feature.

Do you mean the crash detection? If it is, then that's great!

> However, Lars said that it should be pushed only to 6.1.

WL#4621 or the crash detection?

If crash detection is not hard to implement and is risk free, I think we
can push it to 6.0.

> It is up to you. Your "wishes is my command".

I think you mean Lars here, :-)

> >   
> >> - The better recovery mechanism (WL#4621) can be implemented later.
> >>
> >> /L
> >>
> >>
> >>     
> >>> Alfranio Correia wrote:
> >>>       
> >>>> #At
> file:///home/acorreia/workspace.sun/playground.mysql/mysql-5.0.68-custom_build/
> >>>>
> >>>>  2705 Alfranio Correia	2008-10-29
> >>>>       Introduced a recovery mechanism that throws away
> relay-log.bin* retrieved
> >>>>       from a master in case of a crash and updates the master.info
> based on
> >>>>       the relay.info. This is done because such files may be
> corrupted after
> >>>>       a crash. We are assuming that relay.info is fsynced per
> transaction.
> >>>>       See BUG#40337.
> >>>>         
> >
> >   
> 
> Cheers.
>