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
