Skipping. 8.11.1 By mutual agreement of the parties, any step of the grievance procedure may be omitted. At its discretion, the Association may initiate an Association grievance at Step 2
Skipping. Since the execution replicas can only execute requests in the specified total order, it is important that a single slow cluster does not delay the other clusters. To avoid these delays, when a cluster sends a batch of requests to the execution cluster, it also sends a ⟨FLUSH, n⟩ to ⟨ ⟩ all other clusters, where n is the sequence number of the slot being filled. Upon receipt of such a message, the leader of each cluster “flushes” any sequence num- ber that is smaller than n by proposing no-ops for those slots. Specifically, it sends a SKIP, n[ ] to all execution replicas, where n[ ] is an array of sequence numbers that it skips (i.e. proposes a no-op for). This mechanism is similar in spirit to skipping in Mencius [16]. Note that it is not necessary to send the FLUSH mes- sage to all replicas of a cluster. Instead, we can send it only to the leader. Even if the leader is crashed, the sending cluster eventually learns who the new leader is (Section 3.4) and immediately sends a FLUSH message to that replica, thus flushing all previous slots.