Can't schedule 4 TiKV instances on 3 K8s nodes

[weekface]

We have a new algorithm here, there is a corner that we can't schedule 4 tikv instances on a 3 k8s nodes.

tidb-operator/pkg/scheduler/predicates/ha.go

Lines 135 to 137 in 93d5c06

	if podsCount+1 >= int(replicas+1)/2 {
	continue
	}

tidb-operator/pkg/scheduler/predicates/ha_test.go

Lines 754 to 773 in 93d5c06

	{
	name: "three nodes, three pods scheduled on these three nodes, replicas is 4, can't scheduled",
	podFn: newHAPDPod,
	nodesFn: fakeThreeNodes,
	podListFn: podListFn(map[string][]int32{"kube-node-1": {0}, "kube-node-2": {1}, "kube-node-3": {2}}),
	acquireLockFn: acquireSuccess,
	tcGetFn: func(ns string, tcName string) (*v1alpha1.TidbCluster, error) {
	tc, _ := tcGetFn(ns, tcName)
	tc.Spec.PD.Replicas = 4
	return tc, nil
	},
	expectFn: func(nodes []apiv1.Node, err error, recorder record.FakeRecorder) {
	g.Expect(err).To(HaveOccurred())
	events := collectEvents(recorder.Events)
	g.Expect(events).To(HaveLen(1))
	g.Expect(events[0]).To(ContainSubstring("FailedScheduling"))
	g.Expect(strings.Contains(err.Error(), "can't schedule to nodes:")).To(BeTrue())
	g.Expect(len(nodes)).To(Equal(0))
	},
	},

[cofyc]

what's the expected behavior of HA algorithm?