Currently, Templeton launcher kills all child jobs (jobs tagged with the parent job's id) upon task retry.
Upon templeton launcher task retry, templeton should reconnect to the running job and continue tracking its progress that way.
This logic cannot be used for all job kinds (e.g. for jobs that are driven by the client side like regular hive). However, for MapReduceV2, and possibly Tez and HiveOnTez, this should be the default.