MapReduce programs emit too many metrics

Currently,

• every MapReduce task emits metrics to TMS, each one uses metric tags that include the task id, which becomes part of the row key in the metrics table. If a job has thousands of tasks, then that translates into tens of thousands of write in every iteration of metrics processor
• the MapReduce dirver (MRRuntimeService) periodically fetches the task report and the job counters and emits metrics for all tasks. This can translate into 10s of thousands of metrics in a single message to TMS, which in turn can exceed the message limit (of 10MB by default).

We should:

• evaluate whether any of these metrics are redundant and remove them if so
• evaluate whether any of these metrics are unneeded and remove them if so
• have a way to disable (by config/preference) the task-level metrics (instead only record by task type) to avoid the many writes to the metrics table
• have a way to limit the metric emission from the driver (by config/preference)