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

Switch to Threaded View
Kafka, mail # user - Differences in size of data replicated by mirror maker


Copy link to this message
-
Re: Differences in size of data replicated by mirror maker
Guozhang Wang 2013-08-23, 16:24
We are currently working on the following JIRA to avoid decompress and
re-compress at MirrorMaker, when this is done, the size of the logs on
source and target clusters should be the same as long as the batch size of
the mirror maker producer is the same as the batch size of the source
producer:

https://issues.apache.org/jira/browse/KAFKA-1011
On Fri, Aug 23, 2013 at 7:22 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:

> Ah, one thing to be aware of is that the effectiveness of compression is
> directly related to the producer batch size--more batching, more
> compression. So even if you use compression on both clusters the mirror may
> be much smaller.
>
> -jay
>
> On Friday, August 23, 2013, Rajasekar Elango wrote:
>
> > Thanks Guazhang, Jun,
> >
> > Yes we doing gzip compression and that should be reason for difference in
> > disk usage. I had a typo that the size is actually 91G in source
> cluster.So
> > 25G/91G ratio makes sense for compression.
> >
> > Thanks,
> > Raja.
> >
> >
> > On Thu, Aug 22, 2013 at 7:00 PM, Guozhang Wang <[EMAIL PROTECTED]
> <javascript:;>>
> > wrote:
> >
> > > When you state the numbers, are they the same across instances in the
> > > cluster, meaning that Topic-0 would have 910*5 GB in source cluster and
> > > 25*5 GB in target cluster?
> > >
> > > Another possibility is that MirrorMaker uses compression on the
> producer
> > > side, but I would be surprised if the compression rate could be 25/910.
> > >
> > > Guozhang
> > >
> > >
> > > On Thu, Aug 22, 2013 at 3:48 PM, Rajasekar Elango <
> > [EMAIL PROTECTED] <javascript:;>
> > > >wrote:
> > >
> > > > Yes, both source and target clusters have 5 brokers in cluster.
> > > >
> > > > Sent from my iPhone
> > > >
> > > > On Aug 22, 2013, at 6:11 PM, Guozhang Wang <[EMAIL PROTECTED]
> <javascript:;>>
> > wrote:
> > > >
> > > > > Hello Rajasekar,
> > > > >
> > > > > Are the size of the source cluster and target cluster the same?
> > > > >
> > > > > Guozhang
> > > > >
> > > > >
> > > > > On Thu, Aug 22, 2013 at 2:14 PM, Rajasekar Elango <
> > > > [EMAIL PROTECTED] <javascript:;>>wrote:
> > > > >
> > > > >> Hi,
> > > > >>
> > > > >> We are using mirrormaker to replicate data between two kafka
> > clusters.
> > > > I am
> > > > >> seeing huge difference in size of log in data dir between the
> broker
> > > in
> > > > >> source cluster vs broker in destination cluster:
> > > > >>
> > > > >> For eg: Size of ~/data/Topic-0/ is about 910 G in source broker,
> but
> > > > only
> > > > >> its only 25G in destination broker. I see segmented log files
> (~500
> > M)
> > > > is
> > > > >> created for about every 2 or 3 mins in source brokers, but I see
> > > > segmented
> > > > >> log files is created for about every 25 mins in destination
> broker.
> > > > >>
> > > > >> I verified mirrormaker is doing fine using consumer offset
> checker,
> > > not
> > > > >> much lag, offsets are incrementing. I also verified that
> > > > topics/partitions
> > > > >> are not under replicated in both source and target cluster. What
> is
> > > the
> > > > >> reason for this difference in disk usage?
> > > > >>
> > > > >>
> > > > >> --
> > > > >> Thanks,
> > > > >> Raja.
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > -- Guozhang
> > > >
> > >
> > >
> > >
> > > --
> > > -- Guozhang
> > >
> >
> >
> >
> > --
> > Thanks,
> > Raja.
> >
>

--
-- Guozhang