statistics.histogram_statistics.lower_bound = 0.0
    statistics.histogram_statistics.bin_width = 1.0

    hist_freq = np.zeros(501, dtype=np.int32)

    # Advanced calibration methods that use histograms detect outliers, so they
    # don't use the outliers as min/max values.
    hist_freq[0] = 1
    hist_freq[-1] = 1

    # The majority of the data exists around the center.
    hist_freq[250] = 1000

//     the entity name itself contains a '/', that's ok: "cmd/go/flag" is fine.
//
//   - Words should be '-' separated, as in "gopls/completion/errors-latency"
//
//   - Histograms should use bucket names identifying upper bounds with '<'.
//     For example given two counters "gopls/completion/latency:<50ms" and
//     "gopls/completion/latency:<100ms", the "<100ms" bucket counts events

// CKECK-SAME: <{calibration_method = 5 : i32, id = "skipping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)
// CHECK: %[[CUSTOM_AGGREGATOR_1:.*]], %[[MIN_1:.*]], %[[MAX_1:.*]], %[[HISTOGRAM_1:.*]] = "tf.CustomAggregator"

// timeHistogram represents a distribution of durations in
// nanoseconds.
//
// The accuracy and range of the histogram is defined by the
// timeHistSubBucketBits and timeHistNumBuckets constants.
//
// It is an HDR histogram with exponentially-distributed
// buckets and linearly distributed sub-buckets.
//
// The histogram is safe for concurrent reads and writes.
type timeHistogram struct {

	"math"
	"strings"
	"time"
)

// TimeHistogram is an high-dynamic-range histogram for durations.
type TimeHistogram struct {
	Count                int
	Buckets              []int
	MinBucket, MaxBucket int
}

// Five buckets for every power of 10.
var logDiv = math.Log(math.Pow(10, 1.0/5))

// Add adds a single sample to the histogram.
func (h *TimeHistogram) Add(d time.Duration) {
	var bucket int
	if d > 0 {

type Float64Histogram struct {
	// Counts contains the weights for each histogram bucket.
	//
	// Given N buckets, Count[n] is the weight of the range
	// [bucket[n], bucket[n+1]), for 0 <= n < N.
	Counts []uint64

	// Buckets contains the boundaries of the histogram buckets, in increasing order.
	//
	// Buckets[0] is the inclusive lower bound of the minimum bucket while

apiVersion: release-notes/v2
kind: feature
area: telemetry
issue: []
releaseNotes:
  - |

}

void CalibrationStatisticsCollectorHistogram::Collect(
    const float min, const float max, absl::Span<const int64_t> histogram) {
  if (histogram.empty()) return;

  // Reconstruct the bin width, lower and upper bound from the collected data.
  const float collected_bin_width =
      CalculateBinWidth(min, max, histogram.size());
  const float collected_lower_bound =
      CalculateLowerBound(min, collected_bin_width);

      self.assertLen(aggregator_output.histogram, 512)
      self.assertEqual(sum(aggregator_output.histogram), 5)
      self.assertEqual(aggregator_output.histogram[0], 2)
      self.assertEqual(aggregator_output.histogram[128], 1)
      self.assertEqual(aggregator_output.histogram[192], 1)
      self.assertEqual(aggregator_output.histogram[320], 1)


if __name__ == '__main__':

  auto values_array = static_cast<T*>(TF_TensorData(safe_values_ptr.get()));
  tensorflow::histogram::Histogram histo;
  for (int64_t i = 0; i < TF_TensorElementCount(safe_values_ptr.get()); ++i) {
    const double double_val = static_cast<double>(values_array[i]);
    if (Eigen::numext::isnan(double_val)) {
      std::ostringstream err;
      err << "Nan in summary histogram for: " << k->op_node_name;