maria-developers team mailing list archive

[Jesper Staun Hansen <jesper@xxxxxxxxxxxxxx>]

Hello Gordan,

What does your system variables look like?
What I am wondering about myself is why you're even closing the tables.

Jesper

On 20-03-2013 14:25, Gordan Bobic wrote:
> Just as an additional data point, the query does seem to eventually
> complete (and the transaction it is in commits) after about 2 hours.
> But 2 hours seems an excessively long time for "closing tables". It
> is a simple INSERT ... SELECT ... ON DUPLICATE KEY UPDATE, with a few
> triggers on the table being inserted into firing off. The SELECT on
> it's own takes only seconds to complete. The UPDATEs fired off by the
> triggers complete in milliseconds.
>
> Gordan
>
> On Wed, 20 Mar 2013 12:51:50 +0000, Gordan Bobic <gordan@xxxxxxxxxx> 
> wrote:
> > On Wed, 20 Mar 2013 13:20:28 +0100, Sergei Golubchik
> > <serg@xxxxxxxxxxxx> wrote:
> > > Hi, Gordan!
> > >
> > > On Mar 20, Gordan Bobic wrote:
> > > > > On Mar 20, Gordan Bobic wrote:
> > > > >>  I have a situation with MariaDB 10.0.0 where an INSERT INTO ...
> > > > >>  SELECT query that takes 3 seconds to run ends up stuck in the
> > > > >>  closing tables state for 15+ minutes while burning 100% of CPU.
> > > > >>
> > > >  Stack trace attached in a text file.
> > > >  Does this provide enough information for a worthwhile bug report?
> > >
> > > Was this stack trace created when you had your problem with a stuck
> > > query?
> >
> > Yes. And the same query is still stuck in the same state 35 minutes 
> > later.
> >
> > I have attached a fresh stack trace.
> >
> > > Because it doesn't look like there's any thread stuck in the "closing
> > > table" state.
> >
> > In which case the state shown in "show full processlist" is incorrect.
> > Note that I think that helps.
> >
> > > There's one event running, a couple of idle connection threads, a 
> > > couple
> > > of dead connection thread (waiting in the thread cache for 
> > > resurrection),
> > > and internal service threads, aria, innodb, signal thread, etc.
> >
> > The one event that is running must be it, then. I specifically disabled
> > everything else on the box (slaves, event schduler) to get a clean 
> > trace.
> >
> > Gordan