## Before Upgrading

Consider the following changes, limitations, and guidelines before you upgrade:

### **API Machinery**

* Ensure the GCI metadata files do not have newline at the end ([#38727](https://github.com/kubernetes/kubernetes/pull/38727), [@Amey-D](https://github.com/Amey-D))
* Significantly speed-up make ([#38700](https://github.com/kubernetes/kubernetes/pull/38700), [@sttts](https://github.com/sttts))

- Number of errors reported to the metric `storage_operation_duration_seconds_count` for emptyDir decreased significantly because previously one error was reported for each projected volume created. ([#117022](https://github.com/kubernetes/kubernetes/pull/117022), [@mpatlasov](https://github.com/mpatlasov)) [SIG Storage]

    /**
     * This type is syntactically identical to "multipart/mixed", but the semantics are different.
     * In particular, in a parallel entity, the order of body parts is not significant.
     */
    @JvmField
    val PARALLEL = "multipart/parallel".toMediaType()

    /**
     * The media-type multipart/form-data follows the rules of all multipart MIME data streams as

// As of go1.22 the compiler supports one special case of a string constant
// being passed to MethodByName. For enterprise customers or those building
// large binaries, this gives a significant reduction in binary size.
// https://github.com/golang/go/issues/62257
func callBackToMethodValue(modelType reflect.Value, cbType callbackType) reflect.Value {
	switch cbType {
	case callbackTypeBeforeCreate:

* Use KubeDNS by default in GCE setups, as CoreDNS has significantly higher memory usage in large clusters. ([#68629](https://github.com/kubernetes/kubernetes/pull/68629), [@shyamjvs](https://github.com/shyamjvs))

functionality to complete writing data to the datastore. Object storage operations are atomic by nature and they do not require/implement rename API. The default S3A committer emulates renames through copy and delete APIs. This interaction pattern causes significant loss of performance because of the write amplification. _Netflix_, for example, developed two new staging committers - the Directory staging committer and the Partitioned staging committer - to take full advantage of native object storage operations....

### Bug or Regression

- An inefficient lock in EndpointSlice controller metrics cache has been reworked. Network programming latency may be significantly reduced in certain scenarios, especially in clusters with a large number of Services. ([#107167](https://github.com/kubernetes/kubernetes/pull/107167), [@robscott](https://github.com/robscott)) [SIG Apps and Network]

    checkpoints continue to be readable; controlled by the `write_version`
    argument, `tf.train.Saver` now by default writes out in the new V2 format.
    It significantly reduces the peak memory required and latency incurred
    during restore.
*   Added a new library for library of matrix-free (iterative) solvers for

### Bug or Regression

- An inefficient lock in EndpointSlice controller metrics cache has been reworked. Network programming latency may be significantly reduced in certain scenarios, especially in clusters with a large number of Services. ([#107168](https://github.com/kubernetes/kubernetes/pull/107168), [@robscott](https://github.com/robscott)) [SIG Apps and Network]