Re: what is fat client and FatClientTimeout?
> I don't know what a Fat Client is. Got from google.. http://svn.apache.org/repos/asf/cassandra/trunk/examples/client_only/README.txt http://www.datastax.com/dev/blog/improved-hadoop-output > And why are they removed automatically from gossip after FatClientTimeout which is half QUARANTINE_TIME? What's the concerns behind it? Dig into the code, the comment explain it. // half of QUARATINE_DELAY, to ensure justRemovedEndpoints has enough leeway to prevent re-gossip FatClientTimeout = (long)(QUARANTINE_DELAY / 2); On Sat, Jan 5, 2013 at 9:41 AM, Manu Zhang wrote: > Hi, I'm going through the Gossip codes of Cassandra-1.2-snapshot. > Since I started to follow Cassandra recently, I don't know what a Fat > Client is. And why are they removed automatically from gossip after > FatClientTimeout which is half QUARANTINE_TIME? What's the concerns behind > it? > > > Thanks. > > Manu Zhang >
decommission and hints
Hi all, Looks like decommissioning a node (nodetool decommission) will stream all the non-system table data out to it's appropriate peers (StorageService.unbootstrap()), but what should happen to hints? From reading StorageService, it seems like nothing in particular forces them to get replayed (which makes sense) or stashed onto another, live node. The use case here is auto-scaling vnode clusters in ec2. In the normal usage (that is, today in c* 1.1 running in the cloud), we never call decommission as clusters always remain at a fixed, stable size. However, when auto-scaling down a vnode cluster, I'd want to call decommission on the node to leave the ring (and be terminated). I'm worried about terminating a node and it's hints never getting played before it disappears. Node repair, of course, would help on the anti-entropy front, but wondering if there's value in streaming the hints to a peer (in the same rack?). Thanks, -Jason Brown Netflix
