Hello Cassandra Devs, We are currently trying to optimize our Cassandra system with different workloads. One of our workload is update heavy (very).
Currently we are running with a patch that allows the Live Ratio to go below 1.0 (lower bound set to 0.1 now) which gives us a bit better performance in terms of flushes on this particular CF. We then experienced unexpected memory issues which on further inspection seems to be related to the SlabAllocator. What happens is that we allocate a Region of 1MB every couple of seconds (the columns we write in this CF contain serialized session data, can be 100K each), so overwrites are actually done into another Region and these regions are only freed (most of the time) when the Memtable is flushed. We actually added some debug logs and to write about 300MB to disk we created roughly 3000 regions. (3GB of data, some of them might be collected before the flush but probably not much) It would really great if we could use the native allocator only for this CF. Since the SlabAllocator gives us very good results on our other CFs. (we tried running on a patched version with the HeapAllocator set but went OOM almost immediately) I have found this issue in which Jonathan mentions he is ok with adding a configuration option: https://issues.apache.org/jira/browse/CASSANDRA-3073 Unfortunately it seems the issue was closed and nothing was implemented. Would you guys consider adding this option to a future release? SlabAllocator should be the default but in the CF properties the HeapAllocator can be set. If you want I can try to create a patch myself and submit it to you? Kind Regards Joost -- Joost van de Wijgerd Visseringstraat 21B 1051KH Amsterdam +31624111401 joost.van.de.wijgerd@Skype http://www.linkedin.com/in/jwijgerd
