Are you writing these records as batches (with begin batch, not begin unlogged)? As a logged batch ('begin batch' is logged) gets written to a system keyspace with CL before being applied to the underlying CF, and I'm not sure we want to go special casing that for non-durable writes, since you probably should just be using unlogged batches if you don't care.



On 27 March 2014 15:28, Keith Freeman <8forty@gmail.com> wrote:
We're running an insert-heavy use-case and have set "durable_writes = false" for all of our keyspaces.  While inserts are coming in (about 2000 1k-records per second), we are still seeing 50Mb written to files in the commitlog directory every 6-10 seconds (using iostat).

Anybody know why so much data would written in this situation? We're guessing that these writes are related to the system and system_traces keyspaces, right?  Is it considered ok to turn off durable_writes on those keyspaces?  (I couldn't find anything on this searching the message board or google).

(Note we have double- and triple- checked that all of our keyspaces definitely have durable_writes=false, with multiple clean-and-restarts of cassandra)