If I was rebuilding my power after spending the first thousand years of the Third Age as a shapeless evil I would cast my Eye of Fire in the direction of the filthy little multi_gets.
A node can fail to respond to a query with rpc_timeout for two reasons: either the command did not run or the command started but did not complete. The former is much more likely. If it is happening you will see large pending counts and dropped messages in nodetool tpstats, you will also see log entries about dropped messages.
When you send a multi_get each row you request becomes a message in the read thread pool. If you request 100 rows you will put 100 messages in the pool, which by default has 32 threads. If some clients are sending large multi get (or batch mutations) you can overload nodes and starve other clients.
Hope that helps.
On 6/05/2012, at 7:14 AM, Luís Ferreira wrote:
I'm doing get_slice on huge rows (3 million columns) and even though I am doing it iteratively I keep getting TimeoutExceptions. I've tried to change the number of columns fetched but it did not work.
I have a 5 machine cluster, each with 4GB of which 3 are dedicated to cassandra's heap, but still the all consume all of the memory and get huge IO wait due to the amout of reads.
I am running tests with 100 clients all performing multiple operations mostly get_slice, get and multi_get, but the timeouts only occur in the get_slice.
Does this have anything to do with cassandra's ability (or lack thereof) to keep the rows in memory? Or am I doing anything wrong? Any tips?