Today, we looked and noticed that the hdfs-balancer had not been successfully running for an unknown period of time, due to an 'unfinalized' HDFS upgrade. We finalized it (not sure why this was not finalized), and balancer ran fine.

We should at least monitor that hdfs-balancer is running and actually doing stuff, or even better, somehow monitor the 'balancedness' of HDFS blocks across the cluster.