Documentation for etcd-io/raft package describes the Asynchronous Storage Writes feature:
...an alternate interface for local storage writes that can provide better performance in the presence of high proposal concurrency by minimizing interference between proposals...
Independent of performance benefits, I find this approach leads to a cleaner separation of code handling "appended" and "committed" entries. We're following the skeleton of the new "state machine handling loop", as provided in the documentation: 09-async-storage-write/the_test.go.
Some changes that were necessary to make this work:
Snapshot proto now comes from a Message, and can be unset (nil), so this requires an explicit check.HardState needs to be reconstructed from fields of the Message (as opposed to being a separate field in Ready).Note:
The test disables the simulation of failed message delivery (introduced in 08-running-cluster, because sometimes a message containing the
ConfChangeentry isn't delivered and the new node never joins the cluster.To re-enable message delivery failures, the test would need to detect that a configuration change proposal was lost and retry.
Next: 10-separate-process.