This issue is marked as Fixed. Open a new JIRA with a description about the issue and attach the patch.
Rajiv, the reason why there is no new jira is because this is a bug that occurs only on the Yahoo! distribution and not on trunk.
To give an explanation, the patch
HADOOP-5420-v20.patch introduced some code to make sure the pid files written by the task-controller were owned by the tasktracker process, as a security check. Inadvertently, in this patch, we removed some code that changed the ownership of the pid file (which was written as root) to be owned by the TT user. As a result, pid files were created as root, but the new check introduced in the patch failed during kill because it found the PID files were not owned by the TT user and hence treated them as suspect. Hence tasks failed to be killed causing runaway processes on the cluster.
The attached patch re-introduces the code that changes ownership of the pid file to the TT user so that during killing the security check would pass and processes would be killed.