.that(percentiles().index(index).compute(PSEUDORANDOM_DATASET))
          .isWithin(ALLOWED_ERROR)
          .of(expectedLargeDatasetPercentile(index));
    }
  }

  @AndroidIncompatible // slow
  public void testPercentiles_index_computeInPlace() {
    // Assert that the computation gives the correct result for all possible percentiles.
    for (int index = 0; index <= 100; index++) {

 *
 * <p>To compute the 99th percentile:
 *
 * <pre>{@code
 * double myPercentile99 = percentiles().index(99).compute(myDataset);
 * }</pre>
 *
 * where {@link #percentiles()} has been statically imported.
 *
 * <p>To compute median and the 90th and 99th percentiles:
 *
 * <pre>{@code
 * Map<Integer, Double> myPercentiles =
 *     percentiles().indexes(50, 90, 99).compute(myDataset);
 * }</pre>
 *

        regA(builder);
        return builder;
    }

    protected PercentilesAggregationBuilder regPercentilesA(String name, String field) {
        PercentilesAggregationBuilder builder = AggregationBuilders.percentiles(name).field(field);
        regA(builder);
        return builder;
    }

    protected PercentileRanksAggregationBuilder regPercentileRanksA(String name, String field, double[] values) {

namespace. It then executes a series of queries per requested workload to
find the following top-level workload metrics: total requests per second,
error rate, and request latency at p50, p90, and p99 percentiles. The
query results are printed to the console, organized by workload name.

All metrics returned are from server-side reports. This means that latencies

        regA(builder);
        return builder;
    }

    protected PercentilesAggregationBuilder regPercentilesA(String name, String field) {
        PercentilesAggregationBuilder builder = AggregationBuilders.percentiles(name).field(field);
        regA(builder);
        return builder;
    }

    protected PercentileRanksAggregationBuilder regPercentileRanksA(String name, String field, double[] values) {

        regA(builder);
        return builder;
    }

    protected PercentilesAggregationBuilder regPercentilesA(String name, String field) {
        PercentilesAggregationBuilder builder = AggregationBuilders.percentiles(name).field(field);
        regA(builder);
        return builder;
    }

    protected PercentileRanksAggregationBuilder regPercentileRanksA(String name, String field, double[] values) {

    }
    return dummy;
  }

  @Benchmark
  double percentile90(int reps) {
    double dummy = 0.0;
    for (int i = 0; i < reps; i++) {
      dummy += algorithm.singleQuantile(90, 100, dataset(i));
    }
    return dummy;
  }

  @Benchmark
  double percentile99(int reps) {
    double dummy = 0.0;
    for (int i = 0; i < reps; i++) {

    }
    return dummy;
  }

  @Benchmark
  double percentile90(int reps) {
    double dummy = 0.0;
    for (int i = 0; i < reps; i++) {
      dummy += algorithm.singleQuantile(90, 100, dataset(i));
    }
    return dummy;
  }

  @Benchmark
  double percentile99(int reps) {
    double dummy = 0.0;
    for (int i = 0; i < reps; i++) {

introduced significant changes to its garbage collection algorithm. As a result, we observed an increase in memory usage for kube-apiserver in larger an heavily loaded clusters up to ~25% (with the benefit of API call latencies drop by up to 10x on 99th percentiles). If the memory increase is not acceptable for you you can mitigate by setting GOGC env variable (for our tests using GOGC=63 brings memory usage back to original value, although the exact value may depend on usage patterns on your cluster). ([...

* The kubelet --eviction-minimum-reclaim option can now take percentages as well as absolute values for resources quantities ([#33392](https://github.com/kubernetes/kubernetes/pull/33392), [@sjenning](https://github.com/sjenning))