"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",
    },

		}
		atomic.StoreInt64(&dt.timeout, timeout)
	} else if failPct < dynamicTimeoutDecreaseThresholdPct {
		// We are hitting the timeout relatively few times,
		// so decrease the timeout towards 25 % of maximum time spent.
		max = max * 125 / 100

		timeout := atomic.LoadInt64(&dt.timeout)
		if max < time.Duration(timeout) {
			// Move 50% toward the max.
			timeout = (int64(max) + timeout) / 2
		}

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.

        }
        return sb.toString();
      }
    },
    // It really seems like this should be faster than TO_BYTE_ARRAY_NEW_STRING.  But it just isn't
    // my best guess is that the jdk authors have spent more time optimizing that callpath than this
    // one. (StringCoding$StringDecoder vs. StreamDecoder).  StringCoding has a ton of special cases

		dur = 0
	}
	a.Total += int64(dur)
	a.N++
}

// Merge b into a.
func (a *AccElem) merge(b AccElem) {
	a.N += b.N
	a.Total += b.Total
	a.Size += b.Size
}

// Avg returns average time spent.
func (a AccElem) avg() time.Duration {
	if a.N >= 1 && a.Total > 0 {
		return time.Duration(a.Total / a.N)
	}
	return 0
}

// asTimedAction returns the element as a madmin.TimedAction.

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();

    return (totalLoadCount == 0) ? 0.0 : (double) loadExceptionCount / totalLoadCount;
  }

  /**
   * Returns the total number of nanoseconds the cache has spent loading new values. This can be
   * used to calculate the miss penalty. This value is increased every time {@code loadSuccessCount}
   * or {@code loadExceptionCount} is incremented.
   */

    return (totalLoadCount == 0) ? 0.0 : (double) loadExceptionCount / totalLoadCount;
  }

  /**
   * Returns the total number of nanoseconds the cache has spent loading new values. This can be
   * used to calculate the miss penalty. This value is increased every time {@code loadSuccessCount}
   * or {@code loadExceptionCount} is incremented.
   */