Hello. I have been looking into the memory patterns of Cassandra (1.1.5), and I have a suggestion that I haven't seen discussed here.
Cassandra is configured by default with a GC tailored for pauseless operation instead of throughput, and this makes sense since it needs to answer client queries with low latency. Under "normal" operation I see no Full GCs. But when repairs/compactions are run the memory pattern changes and I see frequent Full GCs, which take too long. Wouldn't it make sense to do repairs/compacts in a separate JVM process with a GC configured differently? In this case pauses wouldn't pose such a problem. I'm finding it difficult to configure the GC to behave correctly with and without repairs going on. Best regards, André
signature.asc
Description: Message signed with OpenPGP using GPGMail