My experience to debug this kind of issue is to turn on trace. The nice thing in cassandra is: you can turn on trace only on 1 node and with a small percentage, i.e. nodetool settraceprobability 0.05 --- only run on 1 node. Hope it helps.
Regards, James On Thu, Jul 21, 2022 at 2:50 PM Tolbert, Andy <x...@andrewtolbert.com> wrote: > I'd bet the JIRA that Paul is pointing to is likely what's happening > here. I'd look for read repair errors in your system logs or in your > metrics (if you have easy access to them). > > There are operations that can happen during the course of a query > being executed that may happen at different CLs, atomic batch log > timeouts (CL TWO I think?) and read repair came to my mind (especially > for CL ALL) that can make the timeout/unavailable exceptions include a > different CL. I also remember some DSE features causing this as well > (rbac, auditing, graph and solr stuff). In newer versions of C* the > errors may be more specific or a warning may come along with it > depending on what is failing. > > Thanks, > Andy >