autopilot: don't follow the normal server removal rules for nonvoters

agent/consul/autopilot/autopilot.go

@@ -204,15 +204,22 @@ func (a *Autopilot) pruneDeadServers() error {
 		}
 		if server != nil {
 			// todo(kyhavlov): change this to index by UUID
-			if _, ok := staleRaftServers[server.Addr.String()]; ok {
+			s, found := staleRaftServers[server.Addr.String()]
+			if found {
 				delete(staleRaftServers, server.Addr.String())
 			}
 
 			if member.Status == serf.StatusFailed {
+				// If the node is a nonvoter, we can remove it immediately.
+				if found && s.Suffrage == raft.Nonvoter {
+					a.logger.Printf("[INFO] autopilot: Attempting removal of failed server node %q", member.Name)
+					go serfLAN.RemoveFailedNode(member.Name)
+				} else {
 					failed = append(failed, member.Name)
 				}
 			}
 		}
+	}
 
 	// We can bail early if there's nothing to do.
 	removalCount := len(failed) + len(staleRaftServers)

agent/consul/autopilot_test.go

@@ -92,6 +92,30 @@ func testCleanupDeadServer(t *testing.T, raftVersion int) {
 	}
 }
 
+func TestAutopilot_CleanupDeadNonvoter(t *testing.T) {
+	dir1, s1 := testServer(t)
+	defer os.RemoveAll(dir1)
+	defer s1.Shutdown()
+
+	dir2, s2 := testServerDCBootstrap(t, "dc1", false)
+	defer os.RemoveAll(dir2)
+	defer s2.Shutdown()
+
+	testrpc.WaitForLeader(t, s1.RPC, "dc1")
+
+	// Have s2 join and then shut it down immediately before it gets a chance to
+	// be promoted to a voter.
+	joinLAN(t, s2, s1)
+	retry.Run(t, func(r *retry.R) {
+		r.Check(wantRaft([]*Server{s1, s2}))
+	})
+	s2.Shutdown()
+
+	retry.Run(t, func(r *retry.R) {
+		r.Check(wantRaft([]*Server{s1}))
+	})
+}
+
 func TestAutopilot_CleanupDeadServerPeriodic(t *testing.T) {
 	t.Parallel()
 	dir1, s1 := testServerWithConfig(t, func(c *Config) {