Elasicsearch nodes disconnecting

Question

We have an issue where some nodes in a cluster suddenly leaves the cluster without any apparent reason.

We run on Elasticsearch v0.20.6, JVM 7u25. We use unicast discovery.

This is an embedded ES instance, with 7 nodes in a cluster. Nodes 47, 48, 49 and 50 on one location (network), 24, 25 and 26 on another.

The same thing happens after a while every time, the index files are deleted between the tests. One of the 24, 25, 26 nodes suddenly thinks its the master (which again leads to a split-brain scenario - that is ok and I understand why this happens, but the question is why the disconnect is happening.

First, NODE47 is elected master. All other nodes joins, and things runs smooth for a couple of hours or so.

Then suddenly, here is first traces of that something is visibly going wrong, around 19:10:

Node47:
2013-08-14 19:09:49,243 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][transport_client_worker][T#3]{New I/O worker #3}) [local] disconnected from [[local][VbxjXeqGRIyNFzvK-1JCIw][inet[/**NODE24**:8800]]{local=false}], channel closed event
2013-08-14 19:09:54,109 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][transport_client_worker][T#3]{New I/O worker #3}) [local] disconnected from [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}], channel closed event
2013-08-14 19:10:06,008 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][transport_client_worker][T#4]{New I/O worker #4}) [local] disconnected from [[local][da-T28GDRtWgadrkCvxS-w][inet[/**NODE25**:8800]]{local=false}], channel closed event
2013-08-14 19:10:34,253 TRACE [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][generic][T#19]) [local] [node  ] [[local][VbxjXeqGRIyNFzvK-1JCIw][inet[/**NODE24**:8800]]{local=false}] transport disconnected (with verified connect)
2013-08-14 19:10:34,259 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#24]) [local] connected to node [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}]
2013-08-14 19:10:34,259 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#25]) [local] connected to node [[local][da-T28GDRtWgadrkCvxS-w][inet[/**NODE25**:8800]]{local=false}]
2013-08-14 19:10:34,273 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#26]) [local] connected to node [[local][VbxjXeqGRIyNFzvK-1JCIw][inet[/**NODE24**:8800]]{local=false}]
2013-08-14 19:10:34,290 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#27]) [local] disconnected from [[local][VbxjXeqGRIyNFzvK-1JCIw][inet[/**NODE24**:8800]]{local=false}]


Node24:
2013-08-14 19:10:35,167 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#4]{New I/O worker #4}) [local] [master] pinging a master [local][Y01TgbUzRg-JIIpQ7NqlZg][inet[/**NODE47**:8800]]{local=false} but we do not exists on it, act as if its master failure
2013-08-14 19:10:35,170 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#4]{New I/O worker #4}) [local] [master] stopping fault detection against master [[local][Y01TgbUzRg-JIIpQ7NqlZg][inet[/**NODE47**:8800]]{local=false}], reason [master failure, do not exists on master, act as master failure]
2013-08-14 19:10:35,171 INFO  [org.elasticsearch.discovery.zen] (elasticsearch[local][generic][T#1]) [local] master_left [[local][Y01TgbUzRg-JIIpQ7NqlZg][inet[/**NODE47**:8800]]{local=false}], reason [do not exists on master, act as master failure]
2013-08-14 19:10:35,174 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][clusterService#updateTask][T#1]) [local] [master] restarting fault detection against master [[local][JrRrD5Y8R8WHn1ZAkjYNBw][inet[/**NODE45**:8800]]{local=false}], reason [possible elected master since master left (reason = do not exists on master, act as master failure)]
2013-08-14 19:10:35,181 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#1]) [local] disconnected from [[local][Y01TgbUzRg-JIIpQ7NqlZg][inet[/**NODE47**:8800]]{local=false}]
2013-08-14 19:10:36,233 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#4]{New I/O worker #4}) [local] [master] pinging a master [local][JrRrD5Y8R8WHn1ZAkjYNBw][inet[/**NODE45**:8800]]{local=false} that is no longer a master
2013-08-14 19:10:36,235 INFO  [org.elasticsearch.discovery.zen] (elasticsearch[local][generic][T#5]) [local] master_left [[local][JrRrD5Y8R8WHn1ZAkjYNBw][inet[/**NODE45**:8800]]{local=false}], reason [no longer master]
2013-08-14 19:10:36,235 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#4]{New I/O worker #4}) [local] [master] stopping fault detection against master [[local][JrRrD5Y8R8WHn1ZAkjYNBw][inet[/**NODE45**:8800]]{local=false}], reason [master failure, no longer master]
2013-08-14 19:10:36,241 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][clusterService#updateTask][T#1]) [local] [master] restarting fault detection against master [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}], reason [possible elected master since master left (reason = no longer master)]
2013-08-14 19:10:36,245 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#5]) [local] disconnected from [[local][JrRrD5Y8R8WHn1ZAkjYNBw][inet[/**NODE45**:8800]]{local=false}]
2013-08-14 19:10:37,359 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#3]{New I/O worker #3}) [local] [master] pinging a master [local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false} that is no longer a master
2013-08-14 19:10:37,361 INFO  [org.elasticsearch.discovery.zen] (elasticsearch[local][generic][T#10]) [local] master_left [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}], reason [no longer master]
2013-08-14 19:10:37,363 DEBUG [org.elasticsearch.discovery.zen.fd] (elasticsearch[local][transport_client_worker][T#3]{New I/O worker #3}) [local] [master] stopping fault detection against master [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}], reason [master failure, no longer master]
2013-08-14 19:10:37,393 DEBUG [org.elasticsearch.transport.netty] (elasticsearch[local][generic][T#10]) [local] disconnected from [[local][V7FXnZiLR-GVIyZ2DOwV2w][inet[/**NODE26**:8800]]{local=false}]

As far as I can read of the logs; this is whats happening:

19:09:49,243 - a channel closed event is received from NODE24 to NODE47 (Master) and it is disconnected 19:10:34,273 - a connection to NODE24 is done, then 19:10:34,290 - we get a "disconnected" from NODE24 19:10:35,167 - NODE24 pings master (NODE47) but the master does not have NODE24 in its list of nodes, and threats this like a master failure.

All of this happening within a second - alas, no timeouts in work here as I know of. Also, there are no large GC or any slowdown that is measurable in this period or before.

Im at loss; why does this happen? If network issues; what should be tested on the network side?

Answer 1

To answer this myself with the actual reason for the behavior;

A tcp-connection between 2 nodes (while keeping the connection to the other nodes) are disconnected. It could be recreated by using a utility like tcpkill.

The Elasticsearch Zen discovery sadly does not handle errors like this very good, and all sorts of strange outcomes are possible. The node that looses connection to the master will do an election, and may confuse other nodes.