On 05/01/2012 04:16 AM, aaron morton wrote:
Gossip information about a node can stay in the cluster for up to 3
days. How long has this been going on for ?
This has been going for over a week already without any signs of slow
down, all nodes that have changed ip popup as UP/DEAD endlessly.
Any ideas?
Thanks
I'm unsure if this is expected behaviour. But it sounds like Gossip is
kicking out the phantom node correctly.
Can you use nodetool gossipinfo on the nodes to capture some artefacts
while it is still running?
How come the old ip 10.63.14.214 still popup as UP and then declared
as DEAD again, an so on and on?
I think this is gossip bouncing information about the node around.
Once it has been observed as dead for 3 days it should be purged.
Another question, if node is recognised as new (due to ip change) but
with same token - will other nodes stream the hinted handoffs to it?
Hints are stored against the token, not the end point address. When a
node comes up the process is reversed and the end point is mapped to
it's (new) token.
And is there way to tell cassandra also use names and if ip changes
but node name is the same and resolves to the new ip then the cluster
treat it as old node?
Not that I am aware of. It's designed to handle IP addresses changing.
AFAIK the log messages are not indicative of a fault. Instead they
indicate something odd happening with Gossip that is being correctly
handled.
Hope that helps.
-----------------
Aaron Morton
Freelance Developer
@aaronmorton
http://www.thelastpickle.com
On 1/05/2012, at 3:09 AM, Piavlo wrote:
Hi,
We have a cassandra cluster in ec2.
If i stop a node and start it - as a result the node ip changes. The
node is recognised as NEW node and is declared as replacing the
previous node with same token.(But this is the same node of course)
In this specific case the node ip before stop/start was 10.63.14.214
and new ip is 10.54.81.14.
And even that the cluster and node seems to be working fine for more
than a day after the stop/start of this node, I see the following
loop of messages ~ once every minute.
INFO [GossipStage:1] 2012-04-30 14:18:57,089 Gossiper.java (line 838)
Node /10.63.14.214 is now part of the cluster
INFO [GossipStage:1] 2012-04-30 14:18:57,089 Gossiper.java (line 804)
InetAddress /10.63.14.214 is now UP
INFO [GossipStage:1] 2012-04-30 14:18:57,090 StorageService.java
(line 1017) Nodes /10.63.14.214 and cassa1a.internal/10.54.81.14 have
the same token 0. Ignoring /10.63.14.214
INFO [GossipTasks:1] 2012-04-30 14:19:11,834 Gossiper.java (line 818)
InetAddress /10.63.14.214 is now dead.
INFO [GossipTasks:1] 2012-04-30 14:19:27,896 Gossiper.java (line 632)
FatClient /10.63.14.214 has been silent for 30000ms, removing from gossip
INFO [GossipStage:1] 2012-04-30 14:20:30,803 Gossiper.java (line 838)
Node /10.63.14.214 is now part of the cluster
...
How come the old ip 10.63.14.214 still popup as UP and then declared
as DEAD again, an so on and on?
I know since this is ec2 other node with same ip can come UP, but
i've verified and there is no such node and it certainly does not run
cassandra :)
I stop/started another node and observe similar behaviour.
This is version 1.0.8
Another question, if node is recognised as new (due to ip change) but
with same token - will other nodes stream the hinted handoffs to it?
And is there way to tell cassandra also use names and if ip changes
but node name is the same and resolves to the new ip then the cluster
treat it as old node?
Thanks
Alex
