- `apiserver_flowcontrol_demand_seats`: Observations, at the end of every nanosecond, of (the number of seats each priority level could use) / (nominal number of seats for that level)
  - `apiserver_flowcontrol_demand_seats_high_watermark`: High watermark, over last adjustment period, of demand_seats
  - `apiserver_flowcontrol_demand_seats_average`: Time-weighted average, over last adjustment period, of demand_seats

### Misc

- Fixed computing backoff delay when using Job pod failure policy, by including in the backoff delay calculation pod failures ignored from the backoffLimit counter. ([#119434](https://github.com/kubernetes/kubernetes/pull/119434), [@mimowo](https://github.com/mimowo))

- Failed pods associated with a job with `parallelism = 1` are recreated by the job controller honoring exponential backoff delay again. However, for jobs with `parallelism > 1`, pods might be created without exponential backoff delay. ([#115020](https://github.com/kubernetes/kubernetes/pull/115020), [@nikhita](https://github.com/nikhita)) [SIG Apps]

- Fix: use force detach for azure disk ([#91948](https://github.com/kubernetes/kubernetes/pull/91948), [@andyzhangx](https://github.com/andyzhangx)) [SIG Cloud Provider]
- 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

- Fix computing backoff delay when using Job pod failure policy, by including in the backoff delay calculation pod failures ignored from the backoffLimit counter.
  
  Also, compute the backoff delay more accurately for deleted pods. ([#119466](https://github.com/kubernetes/kubernetes/pull/119466), [@mimowo](https://github.com/mimowo)) [SIG Apps]

- Fix: use force detach for azure disk ([#91948](https://github.com/kubernetes/kubernetes/pull/91948), [@andyzhangx](https://github.com/andyzhangx)) [SIG Cloud Provider]
- 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

- Failed pods associated with a job with `parallelism = 1` are recreated by the job controller honoring exponential backoff delay again. However, for jobs with `parallelism > 1`, pods might be created without exponential backoff delay. ([#115021](https://github.com/kubernetes/kubernetes/pull/115021), [@nikhita](https://github.com/nikhita)) [SIG Apps]

* Add cache for VM get operation in azure cloud provider ([#57432](https://github.com/kubernetes/kubernetes/pull/57432), [@karataliu](https://github.com/karataliu))

* Configurable liveness probe initial delays for etcd and kube-apiserver in GCE ([#57749](https://github.com/kubernetes/kubernetes/pull/57749), [@wojtek-t](https://github.com/wojtek-t))

* CRI: `kubectl logs -f` now stops following when container stops, as it did pre-CRI. ([#44406](https://github.com/kubernetes/kubernetes/pull/44406), [@Random-Liu](https://github.com/Random-Liu))
* Azure cloudprovider: Reduce the polling delay for all azure clients to 5 seconds. This should speed up some operations at the cost of some additional quota. ([#43699](https://github.com/kubernetes/kubernetes/pull/43699), [@colemickens](https://github.com/colemickens))