Anthony Milbourne created STORM-2415:
----------------------------------------
Summary: Storm fails to properly handle Zookeeper hosts going down
Key: STORM-2415
URL: https://issues.apache.org/jira/browse/STORM-2415
Project: Apache Storm
Issue Type: Bug
Components: storm-core
Affects Versions: 1.0.3
Environment: All
Reporter: Anthony Milbourne
We run a storm cluster (v.1.0.3) on AWS and have 3 Zookeepers supporting it.
Because AWS sometimes terminates VMs, we sometimes lose a Zookeeper instance.
When this happens, the hostname cannot be resolved for that zookeeper instance
as AWS has taken the VM away. We noticed that in this case storm fails to
connect to zookeeper – even though there are still 2 Zookeeper instances
running. It fails with an exception something like:
{noformat}
java.net.UnknownHostException: zookeeper3
at java.net.InetAddress.getAllByName0(InetAddress.java:1280)
at java.net.InetAddress.getAllByName(InetAddress.java:1192)
at java.net.InetAddress.getAllByName(InetAddress.java:1126)
at
org.apache.storm.shade.org.apache.zookeeper.client.StaticHostProvider.<init>(StaticHostProvider.java:61)
at
org.apache.storm.shade.org.apache.zookeeper.ZooKeeper.<init>(ZooKeeper.java:445)
at
org.apache.storm.shade.org.apache.curator.utils.DefaultZookeeperFactory.newZooKeeper(DefaultZookeeperFactory.java:29)
at
org.apache.storm.shade.org.apache.curator.framework.imps.CuratorFrameworkImpl$2.newZooKeeper(CuratorFrameworkImpl.java:150)
at
org.apache.storm.shade.org.apache.curator.HandleHolder$1.getZooKeeper(HandleHolder.java:94)
at
org.apache.storm.shade.org.apache.curator.HandleHolder.getZooKeeper(HandleHolder.java:55)
at
org.apache.storm.shade.org.apache.curator.ConnectionState.reset(ConnectionState.java:218)
at
org.apache.storm.shade.org.apache.curator.ConnectionState.start(ConnectionState.java:103)
at
org.apache.storm.shade.org.apache.curator.CuratorZookeeperClient.start(CuratorZookeeperClient.java:190)
at
org.apache.storm.shade.org.apache.curator.framework.imps.CuratorFrameworkImpl.start(CuratorFrameworkImpl.java:259)
at org.apache.storm.zookeeper$mk_client.doInvoke(zookeeper.clj:86)
at clojure.lang.RestFn.invoke(RestFn.java:494)
at
org.apache.storm.cluster_state.zookeeper_state_factory$_mkState.invoke(zookeeper_state_factory.clj:28)
at org.apache.storm.cluster_state.zookeeper_state_factory.mkState(Unknown
Source)
<SNIP REST OF STACKTRACE>
{noformat}
Having done some research it looks like this error is caused by a bug in the
Zookeeper client library. There is an issue for it here:
[https://issues.apache.org/jira/browse/ZOOKEEPER-1576]
This issue has been resolved in the version 3.5.x branch of Zookeeper. However,
after 2.5 years and 3 releases the 3.5.x branch of Zookeeper is still in Alpha .
Despite the fact that it is in alpha, there is a branch of Curator (v.3.x.x)
that uses it, but Storm uses Curator version 2.x.x – possibly because it
doesn’t rely on alpha code. So the bug is still unpatched in Storm.
I realise that an upgrade to alpha code may be too much of a risk, but this
problem is a serious issue for those running Storm in a containerised or cloud
environment - so perhaps it may be worth considering?
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
