when
    *   Ensure `tf.distributions.Multinomial` doesn't underflow in `log_prob`.
        Before this change, all partitions of an integer variable were
        initialized with the shape of the unpartitioned variable; after this
        change they are initialized correctly.
*   Other:
    *   Add necessary shape util support for bfloat16.
    *   Add a way to run ops using a step function to MonitoredSession.

- If Openstack LoadBalancer is not defined in cloud config, the loadbalancer will no longer beis not initialized. any more in openstack. All setups must have some setting under that section for the OpenStack provider. ([#65781](https://github.com/kubernetes/kubernetes/pull/65781), [@zetaab](https://github.com/zetaab))

			// Make sure to sleep at least a second to avoid high CPU ticks.
			time.Second)
		time.Sleep(duration)
	}
	c.RLock()
	defer c.RUnlock()
	if c.enabled {
		logger.Info("Cluster replication initialized")
	}
	return nil
}

func (c *SiteReplicationSys) loadFromDisk(ctx context.Context, objAPI ObjectLayer) error {
	buf, err := readConfig(ctx, objAPI, getSRStateFilePath())
	if err != nil {

- Fixed incorrect watch events when a watch is initialized simultanously with a reinitializing watchcache. ([#116436](https://github.com/kubernetes/kubernetes/pull/116436), [@wojtek-t](https://github.com/wojtek-t))

- Fixed a bug that init containers with `Always` restartPolicy may not terminate gracefully if the pod hasn't initialized yet. ([#125935](https://github.com/kubernetes/kubernetes/pull/125935), [@gjkim42](https://github.com/gjkim42)) [SIG Node and Testing]
- Fixed a bug where `kubectl describe` incorrectly displayed NetworkPolicy port ranges

* If CSINodeInfo and CSIMigration feature flags are active in the cluster, Kubelet will post NotReady until CSINode is initialized with basic volume plugin mechanism information for well-known drivers ([#74835](https://github.com/kubernetes/kubernetes/pull/74835), [@davidz627](https://github.com/davidz627))

- For scheduler plugin developers: the scheduler framework's shared PodInformer is now initialized with empty indexers. This enables scheduler plugins to add their extra indexers. Note that only non-conflict indexers are allowed to be added. ([#110663](https://github.com/kubernetes/kubernetes/pull/110663), [@fromanirh](https://github.com/fromanirh))...

## Changelog since v1.17.0-beta.2

### Other notable changes

* kubeadm: fix a panic in case the KubeProxyConfiguration feature gates were not initialized. ([#85524](https://github.com/kubernetes/kubernetes/pull/85524), [@Arvinderpal](https://github.com/Arvinderpal))

- E2e: tests can now register callbacks with `ginkgo.BeforeEach`, `ginkgo.AfterEach` or `ginkgo.DeferCleanup` directly after creating a framework instance and are guaranteed that their code is called after the framework is initialized and before it gets cleaned up. `ginkgo.DeferCleanup` replaces `f.AddAfterEach` and `AddCleanupAction` which got removed to simplify the framework. ([#111998](https://github.com/kubernetes/kubernetes/pull/111998), [@pohly](https://github.com/pohly))...

- Fixed bug where a nonzero exit code was returned when initializing zsh completion even though zsh completion was successfully initialized ([#88165](https://github.com/kubernetes/kubernetes/pull/88165), [@brianpursley](https://github.com/brianpursley)) [SIG CLI]