To test, stop the datanode and the tserver on the same box with SIGSTOP.
After the tserver lock is lost, the master calls recoverLease() on the namenode. The namenode tries to talk to the primary node for the pipeline, which is the datanode which is also stopped. Without these settings, the request takes a long time. Eventually, the master calls recoverLease() again, and that restarts the whole cycle.
There are additional details and analysis in