Personally I think having compression on by default is the wrong choice. Depending on your access patterns and row sizes the overhead of compression can create more Garbage Collection and become your bottleneck before your potentially bottleneck your disk (ssd disk)
On Tue, Feb 18, 2014 at 2:23 PM, Robert Coli <rc...@eventbrite.com> wrote: > On Mon, Feb 17, 2014 at 4:35 PM, Plotnik, Alexey <aplot...@rhonda.ru>wrote: > >> After analyzing Heap I saw this buffer has a size about 70KB per SSTable. >> I have more than 30K SSTables per node. >> > > I'm thinking your problem is not compression, it's using the old 5mb > default for Level Compaction and having 30,000 SSTables as a result. > > Before turning off compression, I suggest : > > 1) change level SSTable size to the new default, 160mb > 2) force all SSTables to L0 (in 1.2, this means removing their .json files > with the node down, IIRC) > 3) watch level compaction run for a long time, reducing the number of > SSTables you have > > As an aside, 1.2.0 beta moved a bunch of data related to compression off > the heap. If you were to try to run the same cluster under 1.1, you'd > probably OOM your heap immediately. > > https://issues.apache.org/jira/browse/CASSANDRA-4941 > > =Rob > >
