Metrics messages are sent on a fixed-delay schedule instead of a fixed-rate schedule

This potentially affects all metrics implementation packages which use a server that expects data to be sent at a particular rate. It means that with counter metrics you will see a spurious periodic dip in the metric, because the rate at which metrics are sent is not quite keeping up with what the server expects.

The fix is trivial. In org.apache.hadoop.metrics.spi.AbstractMetricsContext, the call to Timer.schedule should be changes to Timer.scheduleAtFixedRate (which takes the same arguments).