cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Reik Schatz <reik.sch...@gmail.com>
Subject Re: write count increase after 1.2 update
Date Thu, 17 Jan 2013 10:50:06 GMT
Cool feature, didn't know it existed. It turned however out that everything
works fine! There was a configuration error that duplicated a AWS sns->sqs
subscription, so we go twice the amount of data delivered to our
application. Semi-lame post to this mailing list i guess :( I should have
checked that earlier but major thanks for looking into this and replying.

Cheers,
Reik


On Thu, Jan 17, 2013 at 4:28 AM, aaron morton <aaron@thelastpickle.com>wrote:

> You *may* be seeing this
> https://issues.apache.org/jira/browse/CASSANDRA-2503
>
> It was implemented in 1.1.0 but perhaps data in the original cluster is
> more compacted than the new one.
>
> Are the increases for all CF's are just a few?
> Do you have a work load of infrequent writes to rows followed by wide
> reads ?
>
> Cheers
>
> -----------------
> Aaron Morton
> Freelance Cassandra Developer
> New Zealand
>
> @aaronmorton
> http://www.thelastpickle.com
>
> On 16/01/2013, at 6:23 AM, Reik Schatz <reik.schatz@gmail.com> wrote:
>
> Hi, we are running a 1.1.6 (datastax) test cluster with 6 nodes. After the
> recent 1.2 release we have set up a second cluster - also having 6 nodes
> running 1.2 (datastax).
>
> They are now running in parallel. We noticed an increase in the number of
> writes in our monitoring tool (Datadog). The tool is using the write count
> statistic of nodetool cfstats. So we ran nodetool cfstats on one node in
> each cluster. To get an initial write count. Then we ran it again after 60
> sec. It looks like the 1.2 received about twice the amount of writes.
>
> The way our application is designed is that the writes are idempotent, so
> we don't see a size increase. Were there any changes in between 1.1.6 > 1.2
> that could explain this behavior?
>
> I know that 1.2 has the concept of virtual nodes, to spread out the data
> more evenly. So if the "write count" value was actually the sum of all
> writes to all nodes in the, this increase would make sense.
>
> Reik
>
> ps. the clusters are not 100% identical. i.e. since bloom filters are now
> off-heap, we changed settings for heap size and memtables. Cluster 1.1.6:
> heap 8G, memtables 1/3 of heap. Cluster 1.2.0: heap 4G, memtables 2G. Not
> sure it can have an impact on the problem.
>
>
>

Mime
View raw message