I happned to find this bug. When I am doing some development based on Hadoop-2.6.0. I happend to send to T_KILL signal to one of maptask before this task is register to TaskAttemptListener. Because the attempt is not assigned with a container yet. It may lead to a lot of errors.