"model_path");
auto* load_latency_by_stage = monitoring::Sampler<2>::New(
    {
        "/tensorflow/cc/saved_model/load_latency_by_stage",  // metric name
        "Distribution of wall time spent (in microseconds) in each stage "
        "(restore graph from disk, run init graph op, etc) when loading the "
        "model",
        "model_path",
        "stage",
    },

You can also measure the cumulative time your build agents spent building a changeset, which will give you a sense of the amount of work the CI infrastructure has to exert. The cache's effect here is less money spent on CI resources, as you don't need as many CI agents to maintain the same number of changes built.

using tpu::MlirToHloArgs;
using tpu::ShardingAndIndex;

auto* phase2_bridge_compilation_time = tsl::monitoring::Sampler<1>::New(
    {"/tensorflow/core/tf2xla/api/v2/phase2_compilation_time",
     "The wall-clock time spent on executing graphs in milliseconds.",
     "configuration"},
    // Power of 1.5 with bucket count 45 (> 23 hours)
    {tsl::monitoring::Buckets::Exponential(1, 1.5, 45)});

	"istio.io/istio/pkg/test/util/yml"
)

func (i *istioImpl) Close() error {
	t0 := time.Now()
	scopes.Framework.Infof("=== BEGIN: Cleanup Istio [Suite=%s] ===", i.ctx.Settings().TestID)

	// Write time spent for cleanup and deploy to ARTIFACTS/trace.yaml and logs to allow analyzing test times
	defer func() {
		delta := time.Since(t0)
		i.ctx.RecordTraceEvent("istio-cleanup", delta.Seconds())

            // It isn't clear if storing them uncompressed too would bring a performance benefit,
            // as reading less from the disk may save more time than spent unpacking.
            return compressionMethod != CompressionMethod.STORED;
        }

        private static long computeCrc32Of(byte[] contents) {
            return Hashing.crc32().hashBytes(contents).padToLong();

	processLocksWriteTotalMD          = NewGaugeMD(processLocksWriteTotal, "Number of current WRITE locks on this peer")
	processCPUTotalSecondsMD          = NewCounterMD(processCPUTotalSeconds, "Total user and system CPU time spent in seconds")
	processGoRoutineTotalMD           = NewGaugeMD(processGoRoutineTotal, "Total number of go routines running")

		&compbasemetrics.HistogramOpts{
			Namespace:      namespace,
			Subsystem:      subsystem,
			Name:           "read_wait_seconds",
			Help:           "Histogram of time spent waiting for a watch cache to become fresh.",
			StabilityLevel: compbasemetrics.ALPHA,
			Buckets:        []float64{0.005, 0.025, 0.05, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0, 1.25, 1.5, 2, 3},
		}, []string{"resource"})
)

// maximum of the distribution and false.
//
// Specifically, y is a cumulative duration, and invCumulativeSum
// returns the mutator utilization x such that at least y time has
// been spent with mutator utilization <= x.
func (d *mud) invCumulativeSum(y float64) (float64, bool) {
	if len(d.sorted) == 0 && len(d.unsorted) == 0 {
		return math.NaN(), false
	}

	// Sort edges.
	edges := d.unsorted

	// some lower target to keep the scavenger working.
	memoryLimitGoal atomic.Uint64

	// assistTime is the time spent by the allocator scavenging in the last GC cycle.
	//
	// This is reset once a GC cycle ends.
	assistTime atomic.Int64

	// backgroundTime is the time spent by the background scavenger in the last GC cycle.
	//
	// This is reset once a GC cycle ends.
	backgroundTime atomic.Int64
}

=== Loading projects

It normally doesn’t take a significant amount of time to load projects, nor do you have any control over it.