PD pod unschedulable with scheduler extender

[tennix]

From #170, when deploying a TiDB cluster in DinD with Local PV, even we specify WaitForFirstConsumer binding mode for local volume. The PV and PVC still may get bound before the pod is scheduled.

Previously we don't have scheduler extender, and all pods can be scheduled. But now when adding scheduler extender, PD pods are not allowed to schedule on the same node. So when PV and PVC are bound before pod scheduling, the pod can actually only be scheduled on the node that PV belongs to. But the HA predicate would refuse to schedule on that node, causing pod unschedulable.

This may be related to local volume scheduling bug which uses affinity and anti-affinity rules. It is fixed in Kubernetes v1.12.

[tennix]

This seems affect all location based persistent volumes scheduling. For example GKE regional persistent disk as in #241

[tlightsky]

i meet this issue in OSX DinD env too

[weekface]

@tlightsky Are you using manifests/local-dind/dind-cluster-v1.10.sh to build your DinD env?

[gregwebs]

@tlightsky what is the K8s version?

[tlightsky]

@weekface yes, i used version v1.0.0-beta.1,
then i switched to master, then clean/up again, this problem disappear

[weekface]

this should be fixed in the newer version.