cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Edward Capriolo <>
Subject Re: better anti OOM
Date Wed, 28 Dec 2011 01:08:26 GMT
I do major companions and I have ran into bloom filters causing oom. One
trick I did was using nodetool to lower the size of row/key caches before
triggering the compact and raising them after companion finished. As
suggested running with spare heap is a very good idea it lowers the chance
of a stop the world garbage collection. Also less free space usually means
more memory fragmentation and causes your system to work harder CPU.

it is counter intuitive to leave "free memory"  because you want to get the
large caches etc, but the overhead gives more stability which in the end
gives better performance.

On Tuesday, December 27, 2011, Peter Schuller <>
>>> In general, don't expect to be able to run at close to heap capacity;
>>> there *will* be spikes.
>> i try to tune for 80% of heap.
> Just FYI, at 80% target heap usage you're likely to have fallbacks to
> full compacting GC:s is my guess. If you are doing analytics only and
> aren't latency critical, that's probably fine though.
> --
> / Peter Schuller (@scode,

View raw message