In testing on a cluster, I restarted with one fewer datanodes than previously. This caused a few thousand blocks to be under-replicated. Similar to
HDFS-2795, I saw the under-replicated blocks on the SBN, slowly decreasing as the replication thread ran. This seems to be because we process the replication queue when exiting safemode, even if in standby mode. It also reports many "missing blocks" in the NN UI which are slowly decreasing.