Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Kafka, mail # user - Topic creation on restart


Copy link to this message
-
Re: Topic creation on restart
Neha Narkhede 2013-11-05, 05:12
It is probably due to the same metadata propagation issue.
https://issues.apache.org/jira/browse/KAFKA-1111 should fix that.

Thanks,
Neha
On Mon, Nov 4, 2013 at 11:47 AM, Jason Rosenberg <[EMAIL PROTECTED]> wrote:

> Ok,
>
> After adding a delay before enabling a freshly started broker in the
> metadata vip that clients use, it seems to have drastically reduced the
> number of these topic creation requests.  However, not all of them.
>
> I still see (only some times) a handful of "Topic creation" log messages,
> that happen on a freshly started broker, during the time kafka has started,
> but before the broker has been enabled in the metadata VIP.  Is there an
> explanation for this?
>
> Thanks!
>
> Jason
>
>
> On Fri, Nov 1, 2013 at 4:45 PM, Neha Narkhede <[EMAIL PROTECTED]
> >wrote:
>
> > The mbeans are explained here -
> > http://kafka.apache.org/documentation.html#monitoring. Look for
> > *QueueTimeMs
> >
> > Thanks,
> > Neha
> >
> >
> > On Fri, Nov 1, 2013 at 12:14 PM, Jason Rosenberg <[EMAIL PROTECTED]>
> wrote:
> >
> > > Neha,
> > >
> > > This cluster has on the order of 750 topics.
> > >
> > > It looks like if I add a 20 second delay before placing a broker into
> the
> > > vip for metadata requests, it never seems to have this issue.  So I'm
> not
> > > sure about the 104 seconds number, other than that was how long the
> flood
> > > of "Topic creation" log messages went on for (over 500 of these).
> > >
> > > Which metric should I look at for 'high queue time'?
> > >
> > > Jason
> > >
> > >
> > > On Fri, Nov 1, 2013 at 5:54 AM, Neha Narkhede <[EMAIL PROTECTED]
> > > >wrote:
> > >
> > > > It is proportional to the number of topics but still seems too long.
> > Did
> > > > the broker have high queue time on all requests? Also how many topics
> > > > existed on this cluster?
> > > >
> > > > Thanks,
> > > > Neha
> > > > On Oct 31, 2013 10:56 PM, "Jason Rosenberg" <[EMAIL PROTECTED]>
> wrote:
> > > >
> > > > > In this case, it appears to have gone on for 104 seconds.  Should
> it
> > > take
> > > > > that long?  It doesn't seem to always take this long.  I guess one
> of
> > > the
> > > > > drawbacks of having a large number of topics (brings out the edge
> > > cases).
> > > > >
> > > > > Filed: https://issues.apache.org/jira/browse/KAFKA-1111
> > > > >
> > > > >
> > > > > On Fri, Nov 1, 2013 at 1:17 AM, Neha Narkhede <
> > [EMAIL PROTECTED]
> > > > > >wrote:
> > > > >
> > > > > > >> Would it be possible for the broker to make sure all metadata
> is
> > > > > > propagated, before completing the startup sequence, and taking on
> > > > > metadata
> > > > > > requests?
> > > > > >
> > > > > > Yes, that is the flaw I mentioned before. We need to fix it.
> > > > > >
> > > > > >
> > > > > > On Thu, Oct 31, 2013 at 9:56 PM, Jason Rosenberg <
> [EMAIL PROTECTED]
> > >
> > > > > wrote:
> > > > > >
> > > > > > > Ok,
> > > > > > >
> > > > > > > So, I can safely ignore these, it sounds like.  I don't see any
> > > > > > > corresponding logging around it subsequently not succeeding to
> > > > actually
> > > > > > > create the topic in zk.
> > > > > > >
> > > > > > > Would it be possible for the broker to make sure all metadata
> is
> > > > > > > propagated, before completing the startup sequence, and taking
> on
> > > > > > metadata
> > > > > > > requests?
> > > > > > >
> > > > > > > If I were to not place the newly started broker back into the
> vip
> > > > for a
> > > > > > > delay period after the kafka broker has started, would that
> solve
> > > > this
> > > > > > > issue?
> > > > > > >
> > > > > > > Jason
> > > > > > >
> > > > > > >
> > > > > > > On Fri, Nov 1, 2013 at 12:39 AM, Jun Rao <[EMAIL PROTECTED]>
> > wrote:
> > > > > > >
> > > > > > > > This is probably because some metadata requests were issued
> to
> > a
> > > > > newly
> > > > > > > > started broker before the controller has propagated the
> > metadata.
> > > > > Since
> > > > > > > the
> > > > > > > > broker doesn't have the metadata in the cache, it thinks the