BinaryLogClient race condition on reconnect causing duplicate events to be processed

[andrewbudimanasana]

Timeline:

1. Thread A initial connect
2. Keep-alive thread detects failure when binlog position = X
3. Keep-alive starts reconnect thread B at position X
4. Thread A does not realize that it should stop processing events
5. Thread A and B process duplicate binlogs from position X on up

Currently, the 2 ways thread A was supposed to notice the reconnect and
stop:

1. The socket's input stream gets shut down
2. It processes events only if the global connected state == true

From what I can tell, thread A was supposed to stop because the state
became disconnected, and then exit because the next read should
have yielded EOF.

This doesn't always work because:

1. The input stream is buffered so the EOF read is delayed
2. Before it reaches the end of the buffer, thread B connects
3. Thread A now processes the rest of its buffered events because thread
   B updates the global connect state

The symptom of this is interleaving duplicate events. This especially
messes up the transaction logic.

I'm using mypipe which uses version 0.2.4 so I created a patch on top of that branch here. https://github.com/andrewbudiman/mysql-binlog-connector-java/commits/fix-reconnect-processing

I'd be happy to create a pr for the master branch. I was considering a more intrusive refactor to make things cleaner, but didn't know if you already had ideas so I tried to keep the patch simple.

[shyiko]

@andrewbudimanasana Thank you so much! Fix is going to be released in 0.4.2 (as soon as I'm done testing changes).

[shyiko]

0.4.2 is out (Maven Central sync might take up to 2 hours).