> The spikes in latency don’t seem to be correlated to an increase in reads. > The cluster’s workload is usually handling a maximum workload of 4200 > reads/sec per node, with writes being significantly less, at ~200/sec per > node. Usually it will be fine with this, with read latencies at around 3.5-10 > ms/read, but once or twice an hour the latencies on the 3 nodes will shoot > through the roof. Could there be errant requests coming in from the app ? e.g. something asking for 1’000s of columns ? Or something that hits a row that has a lot of tombstones ?
Take a look at nodetool cfhistograms to see if you have any outlier wide rows. Also the second column, sstables, will tell you how many sstables were touched by reads which. High numbers, above 4, let you know there are some wide rows out there. In 2.0 and later 1.2 releases nodetool cfstats will also include information about the number tombstones touched in a read. Hope that helps. ----------------- Aaron Morton New Zealand @aaronmorton Co-Founder & Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 8/01/2014, at 2:15 am, Jason Wee <peich...@gmail.com> wrote: > /** > * Verbs it's okay to drop if the request has been queued longer than the > request timeout. These > * all correspond to client requests or something triggered by them; we > don't want to > * drop internal messages like bootstrap or repair notifications. > */ > public static final EnumSet<Verb> DROPPABLE_VERBS = > EnumSet.of(Verb.BINARY, > > Verb._TRACE, > > Verb.MUTATION, > > Verb.READ_REPAIR, > Verb.READ, > > Verb.RANGE_SLICE, > > Verb.PAGED_RANGE, > > Verb.REQUEST_RESPONSE); > > > The short term solution would probably increase the timeout in your yaml file > but i suggest you get the monitoring graphs (ping internode, block io) ready > so it will give better indication which might be the exact problem. > > Jason > > > On Tue, Jan 7, 2014 at 2:30 AM, Blake Eggleston <bl...@shift.com> wrote: > That’s a good point. CPU steal time is very low, but I haven’t observed > internode ping times during one of the peaks, I’ll have to check that out. > Another thing I’ve noticed is that cassandra starts dropping read messages > during the spikes, as reported by tpstats. This indicates that there’s too > many queries for cassandra to handle. However, as I mentioned earlier, the > spikes aren’t correlated to an increase in reads. > > On Jan 5, 2014, at 3:28 PM, Blake Eggleston <bl...@shift.com> wrote: > > > Hi, > > > > I’ve been having a problem with 3 neighboring nodes in our cluster having > > their read latencies jump up to 9000ms - 18000ms for a few minutes (as > > reported by opscenter), then come back down. > > > > We’re running a 6 node cluster, on AWS hi1.4xlarge instances, with > > cassandra reading and writing to 2 raided ssds. > > > > I’ve added 2 nodes to the struggling part of the cluster, and aside from > > the latency spikes shifting onto the new nodes, it has had no effect. I > > suspect that a single key that lives on the first stressed node may be > > being read from heavily. > > > > The spikes in latency don’t seem to be correlated to an increase in reads. > > The cluster’s workload is usually handling a maximum workload of 4200 > > reads/sec per node, with writes being significantly less, at ~200/sec per > > node. Usually it will be fine with this, with read latencies at around > > 3.5-10 ms/read, but once or twice an hour the latencies on the 3 nodes will > > shoot through the roof. > > > > The disks aren’t showing serious use, with read and write rates on the ssd > > volume at around 1350 kBps and 3218 kBps, respectively. Each cassandra > > process is maintaining 1000-1100 open connections. GC logs aren’t showing > > any serious gc pauses. > > > > Any ideas on what might be causing this? > > > > Thanks, > > > > Blake > >
