As part of
HDFS-13150, in-progress edit log tailing was changed to use an RPC-based mechanism, thus allowing the edit log tailing frequency to be turned way down, and allowing standby/observer NameNodes to be only a few milliseconds stale as compared to the Active NameNode.
When there is a high volume of transactions on the system, each RPC fetches transactions and takes some time to process them, self-rate-limiting how frequently an RPC is submitted. In a lightly loaded cluster, however, most of these RPCs return an empty set of transactions, consuming a high (de)serialization overhead for very little benefit. This was reported by jojochuang in
HDFS-14276 and I have also seen it on a test cluster where the SbNN was submitting 8000 RPCs per second that returned empty.
I propose we add some sort of backoff to the tailing, so that if an empty response is received, it will wait a longer period of time before submitting a new RPC.