fix: count only the first shutdown ACK from each node so a duplicate cannot stand in for a silent node (#1233)

spinifex/operator/cli/cluster.go

@@ -600,6 +600,8 @@ func drainDHCPLeases(nc *nats.Conn, timeout time.Duration) (released, responders
 // collectShutdownACKs publishes a shutdown request and collects ACKs from nodes.
 // When nodeFilter is non-empty, only ACKs from that node count toward nodeCount
 // and all others are ignored (used by the single-node local-drain path).
+// Only the first ACK from each node counts; repeats are dropped so a duplicate
+// cannot stand in for a node that never answered.
 func collectShutdownACKs(nc *nats.Conn, topic string, reqData []byte, nodeCount int, nodeFilter string, timeout time.Duration) ([]daemon.ShutdownACK, error) {
 	inbox := nats.NewInbox()
 	sub, err := nc.SubscribeSync(inbox)
@@ -614,6 +616,7 @@ func collectShutdownACKs(nc *nats.Conn, topic string, reqData []byte, nodeCount
 	nc.Flush()
 
 	var acks []daemon.ShutdownACK
+	seen := make(map[string]bool, nodeCount)
 	deadline := time.Now().Add(timeout)
 	for len(acks) < nodeCount {
 		remaining := time.Until(deadline)
@@ -631,6 +634,10 @@ func collectShutdownACKs(nc *nats.Conn, topic string, reqData []byte, nodeCount
 		if nodeFilter != "" && ack.Node != nodeFilter {
 			continue
 		}
+		if seen[ack.Node] {
+			continue
+		}
+		seen[ack.Node] = true
 		acks = append(acks, ack)
 	}
 	return acks, nil

spinifex/operator/cli/cluster_test.go

@@ -541,3 +541,39 @@ func TestReportGuestsLeftRunningSevereLogsErrorAndNamesGuests(t *testing.T) {
 	assert.Contains(t, out, "i-abandoned")
 	assert.NotContains(t, out, "i-stopped")
 }
+
+// shutdownACKResponder answers each request on topic with one ACK per entry
+// in nodes, so a repeated name simulates one node replying twice.
+func shutdownACKResponder(t *testing.T, nc *nats.Conn, topic string, nodes ...string) {
+	t.Helper()
+	sub, err := nc.Subscribe(topic, func(msg *nats.Msg) {
+		for _, node := range nodes {
+			body, _ := json.Marshal(daemon.ShutdownACK{Node: node, Phase: "gate"})
+			_ = msg.Respond(body)
+		}
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = sub.Unsubscribe() })
+}
+
+func TestCollectShutdownACKsCountsADuplicateNodeOnce(t *testing.T) {
+	t.Parallel()
+	_, nc, _ := testutil.StartTestJetStream(t)
+	shutdownACKResponder(t, nc, "test.shutdown.dup", "node1", "node1")
+
+	acks, err := collectShutdownACKs(nc, "test.shutdown.dup", []byte(`{}`), 2, "", 300*time.Millisecond)
+	require.NoError(t, err)
+	require.Len(t, acks, 1, "a second ACK from node1 must not stand in for the silent node2")
+	assert.Equal(t, "node1", acks[0].Node)
+}
+
+func TestCollectShutdownACKsNodeFilterIgnoresOtherNodes(t *testing.T) {
+	t.Parallel()
+	_, nc, _ := testutil.StartTestJetStream(t)
+	shutdownACKResponder(t, nc, "test.shutdown.filter", "node2", "node1", "node1")
+
+	acks, err := collectShutdownACKs(nc, "test.shutdown.filter", []byte(`{}`), 1, "node1", 300*time.Millisecond)
+	require.NoError(t, err)
+	require.Len(t, acks, 1)
+	assert.Equal(t, "node1", acks[0].Node)
+}