*  New: Share threads more aggressively between OkHttp's HTTP/2 connections, connection pool,
    web sockets, and cache. OkHttp has a new internal task runner abstraction for managed task
    scheduling. In your debugger you will see new thread names and more use of daemon threads.

 *  Fix: Don't drop callbacks on unexpected exceptions. When an interceptor throws an unchecked

### Action Required

* Add direct serializer ([#26251](https://github.com/kubernetes/kubernetes/pull/26251), [@caesarxuchao](https://github.com/caesarxuchao))
* Add a NodeCondition "NetworkUnavailable" to prevent scheduling onto a node until the routes have been created  ([#26415](https://github.com/kubernetes/kubernetes/pull/26415), [@wojtek-t](https://github.com/wojtek-t))

    // Count how long it actually took to return; we'll accept any number between the expected delay
    // and the approximate actual delay, to be robust to variance in thread scheduling.
    char overWaitNanosFirstDigit =
        Long.toString(thread.timeSpentBlocked - MILLISECONDS.toNanos(longWait ? 5000 : 3000))
            .charAt(0);
    if (overWaitNanosFirstDigit < '4') {

    // Count how long it actually took to return; we'll accept any number between the expected delay
    // and the approximate actual delay, to be robust to variance in thread scheduling.
    char overWaitNanosFirstDigit =
        Long.toString(thread.timeSpentBlocked - MILLISECONDS.toNanos(longWait ? 5000 : 3000))
            .charAt(0);
    if (overWaitNanosFirstDigit < '4') {

  - [stable] Automated docker validation results posted to https://k8s-testgrid.appspot.com/docker [kubernetes/features#57](https://github.com/kubernetes/enhancements/issues/57)
- **Scheduling**

  * New SchedulerExtender enables users to implement custom
out-of-(the-scheduler)-process scheduling predicates and priority functions,
for example to schedule pods based on resources that are not directly managed

- **Scheduling**

	// This benchmark models a situation where spinning in the mutex should be
	// profitable. To achieve this we create a goroutine per-proc.
	// These goroutines access considerable amount of local data so that
	// unnecessary rescheduling is penalized by cache misses.
	m := NewDRWMutex(ds, "")
	var acc0, acc1 uint64
	b.RunParallel(func(pb *testing.PB) {
		var data [16 << 10]uint64
		for i := 0; pb.Next(); i++ {
			m.Lock(id, source)

high/low parts VPERM2I128 $0x20, Y2, Y0, Y4 VPERM2I128 $0x31, Y2, Y0, Y5 VPERM2I128 $0x20, Y3, Y1, Y6 VPERM2I128 $0x31, Y3, Y1, Y7 LEAQ K256<>+0(SB), BP avx2_last_block_enter: ADDQ $0x40, DI MOVQ DI, 520(SP) XORQ SI, SI avx2_loop1: // Do 4 rounds and scheduling VPADDD (BP)(SI*1), Y4, Y9 VMOVDQU Y9, (SP)(SI*1) MOVL AX, DI RORXL $0x19, DX, R13 RORXL $0x0b, DX, R14 ADDL (SP)(SI*1), R11 ORL CX, DI VPALIGNR $0x04, Y6, Y7, Y0 MOVL R9, R15 RORXL $0x0d, AX, R12 XORL R14, R13 XORL R10, R15 VPADDD Y4, Y0, Y0 RORXL...

high/low parts VPERM2I128 $0x20, Y2, Y0, Y4 VPERM2I128 $0x31, Y2, Y0, Y5 VPERM2I128 $0x20, Y3, Y1, Y6 VPERM2I128 $0x31, Y3, Y1, Y7 LEAQ K256<>+0(SB), BP avx2_last_block_enter: ADDQ $0x40, DI MOVQ DI, 520(SP) XORQ SI, SI avx2_loop1: // Do 4 rounds and scheduling VPADDD (BP)(SI*1), Y4, Y9 VMOVDQU Y9, (SP)(SI*1) MOVL AX, DI RORXL $0x19, DX, R13 RORXL $0x0b, DX, R14 ADDL (SP)(SI*1), R11 ORL CX, DI VPALIGNR $0x04, Y6, Y7, Y0 MOVL R9, R15 RORXL $0x0d, AX, R12 XORL R14, R13 XORL R10, R15 VPADDD Y4, Y0, Y0 RORXL...