* Default value for expendable pod priority cutoff in GCP deployment of Cluster Autoscaler changed from 0 to -10.
    * action required: users deploying workloads with priority lower than 0 may want to use priority lower than -10 to avoid triggering scale-up.

### Other notable changes

* Metadata Agent Improvements ([#66485](https://github.com/kubernetes/kubernetes/pull/66485), [@bmoyles0117](https://github.com/bmoyles0117))

## Changelog since v1.15.11

## Changes by Kind

### Bug or Regression

- Client-go: resolves an issue with informers falling back to full list requests when timeouts are encountered, rather than re-establishing a watch. ([#89978](https://github.com/kubernetes/kubernetes/pull/89978), [@liggitt](https://github.com/liggitt)) [SIG API Machinery and Testing]

<pre>
   x        y    min(x, y)    max(x, y)

  -0.0    0.0         -0.0          0.0    // negative zero is smaller than (non-negative) zero
  -Inf      y         -Inf            y    // negative infinity is smaller than any other number
  +Inf      y            y         +Inf    // positive infinity is larger than any other number
   NaN      y          NaN          NaN    // if any argument is a NaN, the result is a NaN
</pre>

- In the event that more than one IngressClass is designated "default", the DefaultIngressClass admission controller will choose one rather than fail. ([#110974](https://github.com/kubernetes/kubernetes/pull/110974), [@kidddddddddddddddddddddd](https://github.com/kidddddddddddddddddddddd)) [SIG Network]

* Fix security holes in GCE metadata proxy. ([#51302](https://github.com/kubernetes/kubernetes/pull/51302), [@ihmccreery](https://github.com/ihmccreery))
* Fixed an issue looking up cronjobs when they existed in more than one API version ([#52227](https://github.com/kubernetes/kubernetes/pull/52227), [@liggitt](https://github.com/liggitt))

* Addon manager supports HA masters. ([#55782](https://github.com/kubernetes/kubernetes/pull/55782), [@x13n](https://github.com/x13n))

  
  vSphere releases less than 67u3 are deprecated as of v1.21. Please consider upgrading vSphere to 67u3 or above. vSphere CSI Driver requires minimum vSphere 67u3.
  

for a specified amount of time. As before, taints are created using `kubectl taint` (but internally they are now represented as a field `taints` in the NodeSpec rather than using the `scheduler.alpha.kubernetes.io/taints` annotation on Node). Tolerations are now specified in a new-in-1.6 `tolerations` field of the PodSpec rather than using the `scheduler.alpha.kubernetes.io/tolerations` annotation on the Pod. When you upgrade to 1.6, the scheduler will start using the fields and will ignore the annotations....

- Fixes a problem with 63-second or 1-second connection delays with some VXLAN-based
  network plugins which was first widely noticed in 1.16 (though some users saw it
  earlier than that, possibly only with specific network plugins). If you were previously
  using ethtool to disable checksum offload on your primary network interface, you should

* Connections from Pods to Services with 0 endpoints will now ICMP reject immediately, rather than blackhole and timeout. ([#72534](https://github.com/kubernetes/kubernetes/pull/72534), [@thockin](https://github.com/thockin))
* Services of type=LoadBalancer which have no endpoints will now immediately ICMP reject connections, rather than time out. ([#74394](https://github.com/kubernetes/kubernetes/pull/74394), [@thockin](https://github.com/thockin))