Fix rolling-update rollback from an unavailable rc

Rolling back from a broken update with only one replica fails with a
timeout in the existing code.

The problem is the scale down logic does not consider unavailable
replicas in the old replication controller when calculating how much to
scale down by. This leads to an obvious problem with a single replica
when min unavailable is 1.

The fix is to allow scaling down all unavailable replicas in the old
controller, while still maintaining the min unavailable invariant.

pkg/kubectl/rolling_updater_test.go

@@ -308,7 +308,7 @@ Scaling foo-v1 down to 0
 				up{8},
 				down{oldReady: 4, newReady: 8, to: 1},
 				up{10},
-				down{oldReady: 10, newReady: 1, to: 0},
+				down{oldReady: 1, newReady: 10, to: 0},
 			},
 			output: `Created foo-v2
 Scaling up foo-v2 from 0 to 10, scaling down foo-v1 from 10 to 0 (keep 9 pods available, don't exceed 12 pods)
@@ -516,7 +516,7 @@ Scaling foo-v1 down to 0
 				up{8},
 				down{oldReady: 4, newReady: 8, to: 2},
 				up{10},
-				down{oldReady: 10, newReady: 2, to: 0},
+				down{oldReady: 1, newReady: 10, to: 0},
 			},
 			output: `Created foo-v2
 Scaling up foo-v2 from 5 to 10, scaling down foo-v1 from 6 to 0 (keep 10 pods available, don't exceed 12 pods)
@@ -542,6 +542,21 @@ Scaling foo-v1 down to 0
 Scaling up foo-v2 from 0 to 20, scaling down foo-v1 from 10 to 0 (keep 10 pods available, don't exceed 70 pods)
 Scaling foo-v2 up to 20
 Scaling foo-v1 down to 0
+`,
+		}, {
+			name:        "1->1 0/1 scale down unavailable rc to a ready rc (rollback)",
+			oldRc:       oldRc(1, 1),
+			newRc:       newRc(1, 1),
+			newRcExists: true,
+			maxUnavail:  intstr.FromInt(0),
+			maxSurge:    intstr.FromInt(1),
+			expected: []interface{}{
+				up{1},
+				down{oldReady: 0, newReady: 1, to: 0},
+			},
+			output: `Continuing update with existing controller foo-v2.
+Scaling up foo-v2 from 1 to 1, scaling down foo-v1 from 1 to 0 (keep 1 pods available, don't exceed 2 pods)
+Scaling foo-v1 down to 0
 `,
 		},
 	}