The problem it is related to the nature how it is resolved the reverse address in ec2 instances.
After isolating the problem I could write a very simple app which reproduces the cause of the issue.
Pass in args the public ip address of the ec2 instance where are you running the following small code.
InetAddress namenodePublicAddress = InetAddress.getByName(args);
System.out.println("getHostAddress: " + namenodePublicAddress.getHostAddress());
System.out.println("getHostName: " + namenodePublicAddress.getHostName());
System.out.println("getCanonicalHostName: " + namenodePublicAddress.getCanonicalHostName());
If I am running it on my laptop I get
if I am running it on ec2 instance
My laptop has the same CentOS 5.5 as my ec2 instance and the /etc/resolv.conf in each cases contains a nameserver entry.
For some unknown reason, the java.net.InetAddress's getHostName() or getCanonicalHostName() does not resolves reverse dns names for ec2 public addresses if it was running in ec2 instance.
But any other resolver tools correctly resolves that reverse dns name.
In whirr codebase there are some getHostName() calls, which because of the previously described symptom, causes that /etc/hadoop/conf/hadoop-site.xml on the worker nodes are incorrectly filled with ip addresses instead of dns names. As we know, it is important to use public dns name of the ec2 instance because amazon's nameserver it can resolve to an external or internal ip address, one that is better for direct communication. In case of hadoop cluster, the used security group does not allow intercommunication between nodes by using public ip address and therefore the worker nodes cannot contact the services on the master node. Looking into the hadoop logs it is clearly visible the problem that workers cannot connect to master.
|Status||Open [ 1 ]||Patch Available [ 10002 ]|
|Component/s||core [ 12313574 ]|
|Assignee||Tibor Kiss [ tibor.kiss ]|
|Fix Version/s||0.3.0 [ 12315487 ]|
|Status||Patch Available [ 10002 ]||Resolved [ 5 ]|
|Resolution||Fixed [ 1 ]|