fix: resolve deadlock when maxSurge>0 rolling update on single-replica LWS

When a LeaderWorkerSet has replicas=1 and maxSurge=1, triggering a
rolling update caused the controller to immediately emit a "deleting
surge replica" event and return replicas=1 (no surge ever created),
leaving the update permanently stuck with the StatefulSet at
partition=1, replicas=1.

Root cause: in Case 2 of rollingUpdateParameters (a new rolling update
is detected) the code called wantReplicas(lwsReplicas). With replicas=1
and maxSurge=1 the condition inside wantReplicas was:

  unreadyReplicas(1) <= maxSurge(1)  →  true

which jumped straight into the "release surge" branch and returned
replicas=1. No surge replica was ever created, so the StatefulSet
partition could never advance.

Fix: Case 2 now returns burstReplicas directly instead of going through
wantReplicas. At the moment a new update is detected all existing
replicas are still running the old template (none are unready due to
the update yet), so the correct action is to expand to
lwsReplicas+maxSurge first. wantReplicas is only meaningful once
stsReplicas==burstReplicas and the surge pods are being replaced.

A new integration test "rolling update with maxSurge=1 and single
replica creates surge before rolling" directly exercises the regression:
it verifies that the leader StatefulSet expands to replicas=2
immediately after the update is triggered, then converges back to
replicas=1 once all groups are ready.

Fixes: #688

Signed-off-by: veast <veast@users.noreply.github.com>

Author: veast

pkg/controllers/leaderworkerset_controller.go

@@ -288,12 +288,15 @@ func (r *LeaderWorkerSetReconciler) rollingUpdateParameters(ctx context.Context,
 	}
 	burstReplicas := lwsReplicas + int32(maxSurge)
 
-	// wantReplicas calculates the final replicas if needed.
+	// wantReplicas calculates the desired total replicas (including any surge)
+	// during an in-progress rolling update based on the number of currently
+	// unready (not-yet-updated) replicas. It is only called after the initial
+	// surge expansion has already happened (i.e. stsReplicas == burstReplicas).
 	wantReplicas := func(unreadyReplicas int32) int32 {
 		if unreadyReplicas <= int32(maxSurge) {
-			// When we have n unready replicas and n bursted replicas, we should
-			// start to release the burst replica gradually for the accommodation of
-			// the unready ones.
+			// We have at most maxSurge unready replicas, meaning all or most of the
+			// surge replicas have been replaced. Release one surge slot per
+			// newly-ready replica so we converge on lwsReplicas.
 			finalReplicas := lwsReplicas + utils.NonZeroValue(int32(unreadyReplicas)-1)
 			r.Record.Eventf(lws, nil, corev1.EventTypeNormal, GroupsProgressing, Delete, fmt.Sprintf("deleting surge replica %s-%d", lws.Name, finalReplicas))
 			return finalReplicas
@@ -302,10 +305,16 @@ func (r *LeaderWorkerSetReconciler) rollingUpdateParameters(ctx context.Context,
 	}
 
 	// Case 2:
-	// Indicates a new rolling update here.
+	// Indicates a new rolling update here. At this point all existing replicas
+	// are still running the old template (none are unready due to the update),
+	// so we must first expand to burstReplicas before the rolling partition
+	// can start advancing. Calling wantReplicas(lwsReplicas) here was wrong:
+	// with replicas=1 and maxSurge=1 it satisfied the shrink condition
+	// (unreadyReplicas(1) <= maxSurge(1)) and immediately returned lwsReplicas,
+	// preventing the surge replica from ever being created.
 	if leaderWorkerSetUpdated {
 		// Processing scaling up/down first prior to rolling update.
-		return min(lwsReplicas, stsReplicas), wantReplicas(lwsReplicas), nil
+		return min(lwsReplicas, stsReplicas), burstReplicas, nil
 	}
 
 	partition := *sts.Spec.UpdateStrategy.RollingUpdate.Partition

test/integration/controllers/leaderworkerset_test.go

@@ -2297,6 +2297,82 @@ var _ = ginkgo.Describe("LeaderWorkerSet controller", func() {
 				},
 			},
 		}),
+
+		ginkgo.Entry("rolling update with maxSurge=1 and single replica creates surge before rolling", &testCase{
+			// Regression test for: with replicas=1 and maxSurge=1, calling
+			// wantReplicas(lwsReplicas) in Case 2 satisfied the shrink condition
+			// (unreadyReplicas(1) <= maxSurge(1)), so the controller returned
+			// replicas=1 and emitted a spurious "deleting surge replica" event
+			// without ever creating the surge pod, leaving the update stuck forever.
+			makeLeaderWorkerSet: func(nsName string) *wrappers.LeaderWorkerSetWrapper {
+				return wrappers.BuildLeaderWorkerSet(nsName).Replica(1).MaxSurge(1)
+			},
+			updates: []*update{
+				{
+					// Set lws to available condition.
+					lwsUpdateFn: func(lws *leaderworkerset.LeaderWorkerSet) {
+						testing.SetSuperPodToReady(ctx, k8sClient, lws, 1)
+					},
+					checkLWSState: func(lws *leaderworkerset.LeaderWorkerSet) {
+						testing.ExpectLeaderWorkerSetAvailable(ctx, k8sClient, lws, "All replicas are ready")
+						testing.ExpectStatefulsetPartitionEqualTo(ctx, k8sClient, lws, 0)
+						testing.ExpectValidLeaderStatefulSet(ctx, k8sClient, lws, 1)
+						testing.ExpectValidWorkerStatefulSets(ctx, lws, k8sClient, true)
+						testing.ExpectLeaderWorkerSetStatusReplicas(ctx, k8sClient, lws, 1, 1)
+					},
+				},
+				{
+					// Trigger rolling update by changing worker template.
+					// The controller must expand to burstReplicas (2) with partition=1
+					// so the surge pod gets the new template. Before the fix, Case 2
+					// returned replicas=1 and the STS never grew.
+					lwsUpdateFn: func(lws *leaderworkerset.LeaderWorkerSet) {
+						testing.UpdateWorkerTemplate(ctx, k8sClient, lws)
+					},
+					checkLWSState: func(lws *leaderworkerset.LeaderWorkerSet) {
+						// replicas must be 2 (1 original + 1 surge) to prove the surge
+						// was actually created. partition=1 means pod-0 (old) is held,
+						// pod-1 (surge, new template) is created first.
+						testing.ExpectValidLeaderStatefulSet(ctx, k8sClient, lws, 2)
+						testing.ExpectLeaderWorkerSetProgressing(ctx, k8sClient, lws, "Replicas are progressing")
+						testing.ExpectLeaderWorkerSetUpgradeInProgress(ctx, k8sClient, lws, "Rolling Upgrade is in progress")
+						testing.ExpectStatefulsetPartitionEqualTo(ctx, k8sClient, lws, 1)
+						testing.ExpectLeaderWorkerSetStatusReplicas(ctx, k8sClient, lws, 1, 0)
+					},
+				},
+				{
+					// Create the surge leader pod and mark it ready. Once pod-1 is
+					// ready the controller releases the surge (replicas=1, partition=0)
+					// and pod-0 starts being replaced with the new template.
+					lwsUpdateFn: func(lws *leaderworkerset.LeaderWorkerSet) {
+						var leaderSts appsv1.StatefulSet
+						testing.GetLeaderStatefulset(ctx, lws, k8sClient, &leaderSts)
+						gomega.Expect(testing.CreateLeaderPods(ctx, leaderSts, k8sClient, lws, 1, 2)).To(gomega.Succeed())
+						testing.SetPodGroupToReady(ctx, k8sClient, lws.Name+"-1", lws)
+					},
+					checkLWSState: func(lws *leaderworkerset.LeaderWorkerSet) {
+						// Surge released: STS back to replicas=1, partition=0.
+						// pod-0 is being replaced (Progressing, not yet Available).
+						testing.ExpectValidLeaderStatefulSet(ctx, k8sClient, lws, 1)
+						testing.ExpectStatefulsetPartitionEqualTo(ctx, k8sClient, lws, 0)
+						testing.ExpectLeaderWorkerSetProgressing(ctx, k8sClient, lws, "Replicas are progressing")
+						testing.ExpectLeaderWorkerSetUpgradeInProgress(ctx, k8sClient, lws, "Rolling Upgrade is in progress")
+					},
+				},
+				{
+					// Mark group-0 ready on the new template; the rollout completes.
+					lwsUpdateFn: func(lws *leaderworkerset.LeaderWorkerSet) {
+						testing.SetPodGroupToReady(ctx, k8sClient, lws.Name+"-0", lws)
+					},
+					checkLWSState: func(lws *leaderworkerset.LeaderWorkerSet) {
+						testing.ExpectLeaderWorkerSetAvailable(ctx, k8sClient, lws, "All replicas are ready")
+						testing.ExpectStatefulsetPartitionEqualTo(ctx, k8sClient, lws, 0)
+						testing.ExpectValidLeaderStatefulSet(ctx, k8sClient, lws, 1)
+						testing.ExpectLeaderWorkerSetStatusReplicas(ctx, k8sClient, lws, 1, 1)
+					},
+				},
+			},
+		}),
 	) // end of DescribeTable
 
 	ginkgo.Context("with gang scheduling enabled", ginkgo.Ordered, func() {