[HADOOP-7469] add a standard handler for socket connection problems which improves diagnostics

• Type: Sub-task
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 0.20.203.0, 0.23.0
• Fix Version/s: 0.23.0, 0.24.0
• Component/s: util
• Labels: debugging
• Target Version/s: 0.23.0, 0.24.0
• Assignee: Steve Loughran
• Reporter: Steve Loughran
• Created: 18/Jul/11 14:30
• Updated: 15/Nov/11 00:50
• Resolved: 30/Sep/11 14:19

Description

connection refused, connection timed out, no route to host, etc, are classic IOExceptions that can be raised in a lot of parts of the code. The standard JDK exceptions are useless for debugging as they

1. don't include the destination (host, port) that can be used in diagnosing service dead/blocked problems
2. don't include any source hostname that can be used to handle routing issues
3. assume the reader understands the TCP stack.
   It's obvious from the -user lists that a lot of people hit these problems and don't know how to fix them. Sometimes the source has been patched to insert the diagnostics, but it may be convenient to have a single method to translate some

   SocketException processIOException(SocketException e, String destHost, int destPort) {
     String localhost = getLocalHostname();
     String details = "From "+ localhost +" to "+ desthost + ":"+destPort;
     if (e instanceof ConnectException) {
       return new ConnectException(details 
               + " -- see http://wiki.apache.org/hadoop/ConnectionRefused --" + e, e);
     }
     if (e instanceof UnknownHostException) {
       return new UnknownHostException(details 
               + " -- see http://wiki.apache.org/hadoop/UnknownHost --" + e, e);
     }
     // + handlers for other common socket exceptions
     
   //and a default that returns an unknown class unchanged
     return e;
   }
     
   

Testing: try to connect to an unknown host, a local port that isn't live, etc. It's hard to replicate all failures consistently. It may be simpler just to verify that if you pass in a specific exception, the string is expanded and the class is unchanged.

This code could then be patched in to places where IO takes place. Note that Http Components and HttpClient libs already add some destination details on some operation failures, with their own HttpException tree: it's simplest to leave these alone.