[
https://issues.apache.org/jira/browse/CASSANDRA-7307?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14018011#comment-14018011
]
Jonathan Ellis commented on CASSANDRA-7307:
-------------------------------------------
Actually I think changing the initial padding (which we may want to do anyway)
is only a band-aid. The real problem is that when a new node processes the
state that is gossiped to it, it treats the heartbeats it gets as if they just
happened, vs (potentially) a long time in the past for a dead node.
I'm not sure it's possible to have it both ways without changing the design: we
want to assume nodes we just learned about are good, to avoid false-positive
failure events a la 6385, but we also want to identify dead nodes quickly.
I think you'd need gossip arrivalwindows for every peer so that you could pass
a "history" to new nodes to fix this. Otherwise, heuristics like "don't count
this node I just learned about if the gossip event is older than X" are going
to be fragile and worse than phi-based FD overall.
> New nodes mark dead nodes as up for 10 minutes
> ----------------------------------------------
>
> Key: CASSANDRA-7307
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7307
> Project: Cassandra
> Issue Type: Bug
> Reporter: Richard Low
> Assignee: Brandon Williams
> Fix For: 1.2.17
>
>
> When doing a node replacement when other nodes are down we see the down nodes
> marked as up for about 10 minutes. This means requests are routed to the dead
> nodes causing timeouts. It also means replacing a node when multiple nodes
> from a replica set is extremely difficult - the node usually tries to stream
> from a dead node and the replacement fails.
> This isn't limited to host replacement. I did a simple test:
> 1. Create a 2 node cluster
> 2. Kill node 2
> 3. Start a 3rd node with a unique token (I used auto_bootstrap=false but I
> don't think this is significant)
> The 3rd node lists node 2 (127.0.0.2) as up for almost 10 minutes:
> {code}
> INFO [main] 2014-05-27 14:28:24,753 CassandraDaemon.java (line 119) Logging
> initialized
> INFO [GossipStage:1] 2014-05-27 14:28:31,492 Gossiper.java (line 843) Node
> /127.0.0.2 is now part of the cluster
> INFO [GossipStage:1] 2014-05-27 14:28:31,495 Gossiper.java (line 809)
> InetAddress /127.0.0.2 is now UP
> INFO [GossipTasks:1] 2014-05-27 14:37:44,526 Gossiper.java (line 823)
> InetAddress /127.0.0.2 is now DOWN
> {code}
> I reproduced on 1.2.15 and 1.2.16.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
