xiangfu0 commented on code in PR #19795: URL: https://github.com/apache/pinot/pull/19795#discussion_r4242056547
########## compatibility-verifier/tdigest-compatibility/GenerateRankErrors.java: ########## @@ -0,0 +1,162 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +import com.tdunning.math.stats.Centroid; +import com.tdunning.math.stats.TDigest; +import java.util.Arrays; +import java.nio.ByteBuffer; +import java.security.MessageDigest; +import java.util.HexFormat; +import java.util.SplittableRandom; +import org.apache.pinot.segment.local.utils.TDigestUtils; + + +/// Independent legacy rank oracle, run sequentially with the pinned baseline utility and t-digest 3.3 only. +public final class GenerateRankErrors { + private static final double[] QUANTILES = {0, 0.5, 0.75, 0.95, 0.99, 1}; + private static final String[] DISTRIBUTIONS = {"UNIFORM", "SKEWED", "HEAVY_TAIL", "BIMODAL", "DUPLICATE_HEAVY"}; + private static final String[] ORDERS = {"ORIGINAL", "REVERSED", "RANDOMIZED"}; + private static final String[] INPUTS = {"SERVER_LOCAL", "DISTRIBUTED"}; + + private GenerateRankErrors() { + } + + public static void main(String[] args) + throws Exception { + System.out.println("# mergeOrders=" + String.join(",", ORDERS)); + System.out.println("# reducerInputs=" + String.join(",", INPUTS)); + for (int fanIn : new int[]{8, 32, 128}) { + for (int compression : new int[]{20, 100, 1000}) { + for (String distribution : DISTRIBUTIONS) { + for (String order : ORDERS) { + for (String input : INPUTS) { + Reference reference = referenceCase(fanIn, compression, distribution, order, input); + System.out.print(fanIn + "/" + compression + "/" + distribution + "/" + order + "/" + input + + "," + reference.inputSha256()); + for (double error : reference.rankErrors()) { + System.out.print("," + error); + } + System.out.println(); + } + } + } + } + } + } + + private record Reference(String inputSha256, double[] rankErrors) { + } + + private static Reference referenceCase(int fanIn, int compression, String distribution, String order, String input) + throws Exception { + TDigest[] sources = new TDigest[fanIn]; + double[] raw = new double[fanIn * 64]; + for (int i = 0; i < fanIn; i++) { + TDigest digest = TDigestUtils.createMergingDigest(compression); + SplittableRandom random = new SplittableRandom(0x5EEDL + i); + for (int j = 0; j < 64; j++) { + double value = value(distribution, random.nextDouble()); + raw[64 * i + j] = value; + digest.add(value); + } + sources[i] = input.equals("DISTRIBUTED") ? TDigestUtils.deserialize(TDigestUtils.serialize(digest)) : digest; + } + TDigest result = TDigestUtils.createMergingDigest(compression); + int[] sourceIndexes = indexes(fanIn, order); + for (int index : sourceIndexes) { + if (result.size() == 0) { + result = sources[index]; + } else { + result.add(sources[index]); + } + } + // Match the final public-compression phase in the historical test before measuring quantiles. + for (Centroid centroid : result.centroids()) { Review Comment: **Oracle measured without the final public-compression pass.** The comment says "Match the final public-compression phase in the historical test before measuring quantiles", but the loop below only validates centroids — it never calls `result.compress()`. In t-digest 3.3, `centroids()`/`quantile()` merge at the internal (~2x) working compression, while `compress()` forces a merge at the public compression. The Pinot side of the comparison (`getReducerRankErrors` in `PercentileTDigestAggregationFunctionTest`) measures after `result.compress()`, so the frozen `tdigest-3.3-k1-rank-errors.csv` envelope is derived from a finer, lower-error digest state than the state being validated. A correct Pinot result that merely matches 3.3's post-compress accuracy can exceed the envelope (padded only by 1/numValues), and any regeneration bakes in the same mismatched baseline. Add `result.compress();` before the quantile measurements. ########## pinot-segment-local/src/main/java/org/apache/pinot/segment/local/customobject/tdigest/NonFiniteAwareTDigest.java: ########## @@ -0,0 +1,665 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package org.apache.pinot.segment.local.customobject.tdigest; + +import java.util.ArrayList; +import java.util.Collection; +import java.util.List; +import javax.annotation.Nullable; +import org.apache.pinot.segment.local.customobject.tdigest.TDigestCodec.SerializedTDigestMetadata; + + +/// Stores non-finite values as tail masses while delegating finite values to an owned native digest. +/// Emits legacy t-digest bytes and preserves historical fractional endpoint provenance across merges and copies. +/// Instances are mutable and require external synchronization. +public final class NonFiniteAwareTDigest extends TDigest { + private PercentileTDigestAccumulator _finiteDigest; + private double _negativeInfinityWeight; + private double _positiveInfinityWeight; + private List<Centroid> _finiteCentroids; + private byte[] _originalFractionalBytes; + private boolean _pendingHistoricalBoundaryRepair; + private double _historicalMin; + private double _historicalMax; + + public static NonFiniteAwareTDigest forLegacyAggregation(double compression) { + return new NonFiniteAwareTDigest(PercentileTDigestAccumulator.forLegacyAggregation(compression), 0.0, 0.0); + } + + private NonFiniteAwareTDigest(PercentileTDigestAccumulator finiteDigest, double negativeInfinityWeight, + double positiveInfinityWeight) { + _finiteDigest = finiteDigest; + _negativeInfinityWeight = negativeInfinityWeight; + _positiveInfinityWeight = positiveInfinityWeight; + } + + /// Snapshots the aggregation without flushing or compressing the source's buffered inputs. + public NonFiniteAwareTDigest copy() { + NonFiniteAwareTDigest copy = + new NonFiniteAwareTDigest(_finiteDigest.copy(), _negativeInfinityWeight, _positiveInfinityWeight); + copy._originalFractionalBytes = _originalFractionalBytes == null ? null : _originalFractionalBytes.clone(); + copy._pendingHistoricalBoundaryRepair = _pendingHistoricalBoundaryRepair; + copy._historicalMin = _historicalMin; + copy._historicalMax = _historicalMax; + return copy; + } + + public static NonFiniteAwareTDigest fromBytes(byte[] bytes) { + SerializedTDigestInput validated = new SerializedTDigestInput(); + validated.reset(bytes); + SerializedTDigestMetadata metadata = validated.getMetadata(); + if (!metadata.hasNonFiniteMeans() || metadata.needsLegacyFallback()) { + return new NonFiniteAwareTDigest(PercentileTDigestAccumulator.fromBytes(validated), 0.0, 0.0); + } + validated.decode(); + if (!PercentileTDigestAccumulator.canNormalizeBoundaries(validated._means, validated._weights, + validated._numCentroids, metadata.min(), metadata.max())) { + return new NonFiniteAwareTDigest(PercentileTDigestAccumulator.fromBytes(validated), 0.0, 0.0); + } + double encodedMin = metadata.min(); + double encodedMax = metadata.max(); + double compression = metadata.compression(); + int centroidCount = validated._numCentroids; + double negativeInfinityWeight = 0.0; + double positiveInfinityWeight = 0.0; + double[] finiteMeans = validated._means; + double[] finiteWeights = validated._weights; + int finiteCount = 0; + double finiteMin = Double.POSITIVE_INFINITY; + double finiteMax = Double.NEGATIVE_INFINITY; + for (int i = 0; i < centroidCount; i++) { + double weight = finiteWeights[i]; + double mean = finiteMeans[i]; + if (weight == 0.0) { + continue; + } + if (mean == Double.NEGATIVE_INFINITY) { + negativeInfinityWeight += weight; + } else if (mean == Double.POSITIVE_INFINITY) { + positiveInfinityWeight += weight; + } else { + finiteWeights[finiteCount] = weight; + finiteMeans[finiteCount] = mean; + finiteMin = Math.min(finiteMin, mean); + finiteMax = Math.max(finiteMax, mean); + finiteCount++; + } + } + + PercentileTDigestAccumulator finiteDigest = + PercentileTDigestAccumulator.forLegacyAggregation(compression); + if (negativeInfinityWeight > 0.0 || positiveInfinityWeight > 0.0) { + finiteDigest.deferBoundaryNormalization(); + } + if (finiteCount > 0) { + finiteMin = Double.isFinite(encodedMin) ? encodedMin : finiteMin; + finiteMax = Double.isFinite(encodedMax) ? encodedMax : finiteMax; + finiteDigest.addHistoricalCentroids(finiteMeans, finiteWeights, finiteCount, finiteMin, finiteMax, true); + } + validated.recordHistoricalFractionalBoundaries(finiteDigest); + NonFiniteAwareTDigest wrapped = + new NonFiniteAwareTDigest(finiteDigest, negativeInfinityWeight, positiveInfinityWeight); + wrapped._pendingHistoricalBoundaryRepair = metadata.weightedBoundaries(); + wrapped._historicalMin = encodedMin; + wrapped._historicalMax = encodedMax; + if (metadata.fractionalWeights()) { + wrapped._originalFractionalBytes = bytes.clone(); + } + return wrapped; + } + + @Override + public void add(double value) { + add(value, 1); + } + + @Override + public void add(double value, double weight) { + if (Double.isNaN(value)) { + throw new IllegalArgumentException("Cannot add NaN to t-digest"); + } + if (!(weight > 0.0) || !Double.isFinite(weight)) { + throw new IllegalArgumentException("TDigest weight must be positive: " + weight); + } + PercentileTDigestAccumulator.checkTotalWeight(getTotalWeight() + weight); + double infinityWeight = 0.0; + if (value == Double.NEGATIVE_INFINITY) { + addInfinityWeight(_negativeInfinityWeight, weight); + infinityWeight = addInfinityWeight(boundaryPreparedView()._negativeInfinityWeight, weight); + } else if (value == Double.POSITIVE_INFINITY) { + addInfinityWeight(_positiveInfinityWeight, weight); + infinityWeight = addInfinityWeight(boundaryPreparedView()._positiveInfinityWeight, weight); + } + requireMutable(); + if (value == Double.NEGATIVE_INFINITY) { + _finiteDigest.deferBoundaryNormalization(); + _finiteDigest.recordFreshFractionalMean(value, weight); + _negativeInfinityWeight = infinityWeight; + } else if (value == Double.POSITIVE_INFINITY) { + _finiteDigest.deferBoundaryNormalization(); + _finiteDigest.recordFreshFractionalMean(value, weight); + _positiveInfinityWeight = infinityWeight; + } else { + _finiteDigest.add(value, weight); + } + mutated(); + } + + @Override + public void add(TDigest other) { + if (other.isEmpty()) { + return; + } + if (!hasValidStatistics()) { + throw PercentileTDigestAccumulator.corruptedMutation(); + } + if (!other.hasValidStatistics() && !isEmpty()) { + throw PercentileTDigestAccumulator.corruptedMutation(); + } + if (other.hasValidStatistics()) { + PercentileTDigestAccumulator.checkTotalWeight(getTotalWeight() + other.getTotalWeight()); + } + if (other instanceof NonFiniteAwareTDigest) { + NonFiniteAwareTDigest wrapped = (NonFiniteAwareTDigest) other; + addInfinityWeight(_negativeInfinityWeight, wrapped._negativeInfinityWeight); + addInfinityWeight(_positiveInfinityWeight, wrapped._positiveInfinityWeight); + NonFiniteAwareTDigest preparedReceiver = boundaryPreparedView(); + NonFiniteAwareTDigest preparedSource = wrapped.boundaryPreparedView(); + double negativeInfinityWeight = + addInfinityWeight(preparedReceiver._negativeInfinityWeight, preparedSource._negativeInfinityWeight); + double positiveInfinityWeight = + addInfinityWeight(preparedReceiver._positiveInfinityWeight, preparedSource._positiveInfinityWeight); + requireMutable(); + wrapped.prepareHistoricalBoundaries(); + if (negativeInfinityWeight > 0.0 || positiveInfinityWeight > 0.0) { + _finiteDigest.deferBoundaryNormalization(); + } + try { + if (!wrapped._finiteDigest.isEmpty()) { + _finiteDigest.add(wrapped._finiteDigest); + wrapped.invalidateDerivedCaches(); + } + } catch (RuntimeException | Error e) { + restoreBoundaryNormalizationAfterFailedMerge(); + throw e; + } + if (wrapped.hasValidStatistics()) { + _finiteDigest.inheritHistoricalFractionalBoundaries(wrapped); + } + _negativeInfinityWeight = negativeInfinityWeight; + _positiveInfinityWeight = positiveInfinityWeight; + mutated(); + return; + } + if (!other.hasValidStatistics() || Double.isFinite(other.getMin()) && Double.isFinite(other.getMax())) { + requireMutable(); + _finiteDigest.add(other); + mutated(); + return; + } + // Read precise centroid weights directly; generic sources need no full digest serialization for their tails. + if (other instanceof PercentileTDigestAccumulator) { + PercentileTDigestAccumulator nativeDigest = (PercentileTDigestAccumulator) other; + if (isEmpty() && nativeDigest.hasUnrepresentableBoundaryPayload()) { + _finiteDigest = nativeDigest.copy(); + mutated(); + return; + } + } + Collection<Centroid> centroids = other.centroids(); + double[] finiteMeans = new double[Math.addExact(centroids.size(), 2)]; + double[] finiteWeights = new double[finiteMeans.length]; + int count = 0; + double sourceMin = other.getMin(); + double sourceMax = other.getMax(); + double negativeInfinityWeight = 0.0; + double positiveInfinityWeight = 0.0; + for (Centroid centroid : centroids) { + double mean = centroid.mean(); + double weight = centroid.weight(); + if (weight == 0.0) { + continue; + } + if (!(weight > 0.0) || !Double.isFinite(weight) || !(mean >= sourceMin && mean <= sourceMax)) { + throw new IllegalArgumentException("Invalid TDigest centroid"); + } + if (mean == Double.NEGATIVE_INFINITY) { + negativeInfinityWeight = other instanceof PercentileTDigestAccumulator + ? negativeInfinityWeight + weight + : addInfinityWeight(negativeInfinityWeight, weight); + } else if (mean == Double.POSITIVE_INFINITY) { + positiveInfinityWeight = other instanceof PercentileTDigestAccumulator + ? positiveInfinityWeight + weight + : addInfinityWeight(positiveInfinityWeight, weight); + } + finiteMeans[count] = mean; + finiteWeights[count++] = weight; + } + // Original contributions exclude unit endpoints introduced by boundary repair. + addInfinityWeight(_negativeInfinityWeight, negativeInfinityWeight); + addInfinityWeight(_positiveInfinityWeight, positiveInfinityWeight); + PercentileTDigestAccumulator.sortCentroids(finiteMeans, finiteWeights, count); + count = PercentileTDigestAccumulator.normalizeSerializedBoundaries( + finiteMeans, finiteWeights, count, sourceMin, sourceMax); + negativeInfinityWeight = 0.0; + positiveInfinityWeight = 0.0; + int finiteCount = 0; + double finiteMin = Double.POSITIVE_INFINITY; + double finiteMax = Double.NEGATIVE_INFINITY; + double totalWeight = getTotalWeight(); + for (int i = 0; i < count; i++) { + double mean = finiteMeans[i]; + double weight = finiteWeights[i]; + totalWeight += weight; + if (mean == Double.NEGATIVE_INFINITY) { + negativeInfinityWeight += weight; + } else if (mean == Double.POSITIVE_INFINITY) { + positiveInfinityWeight += weight; + } else { + finiteMeans[finiteCount] = mean; + finiteWeights[finiteCount++] = weight; + finiteMin = Math.min(finiteMin, mean); + finiteMax = Math.max(finiteMax, mean); + } + } + PercentileTDigestAccumulator.checkTotalWeight(totalWeight); + // Repair may relocate tail mass into a finite header endpoint; merge only the remaining tail mass. + NonFiniteAwareTDigest preparedReceiver = boundaryPreparedView(); + negativeInfinityWeight = addInfinityWeight(preparedReceiver._negativeInfinityWeight, negativeInfinityWeight); + positiveInfinityWeight = addInfinityWeight(preparedReceiver._positiveInfinityWeight, positiveInfinityWeight); + requireMutable(); + if (negativeInfinityWeight > 0.0 || positiveInfinityWeight > 0.0) { + _finiteDigest.deferBoundaryNormalization(); + } + try { + if (finiteCount > 0) { + finiteMin = Double.isFinite(other.getMin()) ? other.getMin() : finiteMin; + finiteMax = Double.isFinite(other.getMax()) ? other.getMax() : finiteMax; + if (other instanceof PercentileTDigestAccumulator) { + _finiteDigest.addHistoricalCentroids(finiteMeans, finiteWeights, finiteCount, finiteMin, finiteMax, false); + } else { + _finiteDigest.addCentroids(finiteMeans, finiteWeights, finiteCount, finiteMin, finiteMax, false); + } + } + } catch (RuntimeException | Error e) { + restoreBoundaryNormalizationAfterFailedMerge(); + throw e; + } + if (!(other instanceof PercentileTDigestAccumulator)) { + for (Centroid centroid : centroids) { + if (Double.isInfinite(centroid.mean())) { + _finiteDigest.recordFreshFractionalMean(centroid.mean(), centroid.weight()); + } + } + } + _finiteDigest.inheritHistoricalFractionalBoundaries(other); + _negativeInfinityWeight = negativeInfinityWeight; + _positiveInfinityWeight = positiveInfinityWeight; + mutated(); + } + + private static double addInfinityWeight(double currentWeight, double addedWeight) { + double totalWeight = currentWeight + addedWeight; + if (currentWeight > 0.0 && addedWeight > 0.0 + && (totalWeight == currentWeight || totalWeight == addedWeight)) { + throw new IllegalArgumentException("TDigest infinity-tail weight addition would absorb positive mass"); + } + return totalWeight; + } + + private void requireMutable() { + if (!hasValidStatistics()) { + throw PercentileTDigestAccumulator.corruptedMutation(); + } + if (_negativeInfinityWeight == 0.0 && _positiveInfinityWeight == 0.0) { + _finiteDigest.prepareForMutation(); + } else { + prepareHistoricalBoundaries(); + } + } + + private void prepareHistoricalBoundaries() { + if (!_pendingHistoricalBoundaryRepair) { + return; + } + double encodedMin = getMin(); + double encodedMax = getMax(); + List<Centroid> centroids = getFiniteCentroids(); + double[] means = new double[Math.addExact(centroids.size(), 4)]; + double[] weights = new double[means.length]; + int count = 0; + if (_negativeInfinityWeight > 0.0) { + means[count] = Double.NEGATIVE_INFINITY; + weights[count++] = _negativeInfinityWeight; + } + for (Centroid centroid : centroids) { + means[count] = centroid.mean(); + weights[count++] = centroid.weight(); + } + if (_positiveInfinityWeight > 0.0) { + means[count] = Double.POSITIVE_INFINITY; + weights[count++] = _positiveInfinityWeight; + } + count = PercentileTDigestAccumulator.normalizeSerializedBoundaries(means, weights, count, encodedMin, encodedMax); + int finiteCount = 0; + double min = Double.POSITIVE_INFINITY; + double max = Double.NEGATIVE_INFINITY; + double negativeInfinityWeight = 0.0; + double positiveInfinityWeight = 0.0; + for (int i = 0; i < count; i++) { + if (means[i] == Double.NEGATIVE_INFINITY) { + negativeInfinityWeight += weights[i]; + } else if (means[i] == Double.POSITIVE_INFINITY) { + positiveInfinityWeight += weights[i]; + } else { + means[finiteCount] = means[i]; + weights[finiteCount++] = weights[i]; + min = Math.min(min, means[i]); + max = Math.max(max, means[i]); + } + } + PercentileTDigestAccumulator finiteDigest = PercentileTDigestAccumulator.forLegacyAggregation(compression()); + finiteDigest.deferBoundaryNormalization(); + min = Double.isFinite(encodedMin) ? encodedMin : min; + max = Double.isFinite(encodedMax) ? encodedMax : max; + finiteDigest.addHistoricalCentroids(means, weights, finiteCount, min, max, true); + finiteDigest.inheritHistoricalFractionalBoundaries(_finiteDigest); + _finiteDigest = finiteDigest; + _negativeInfinityWeight = negativeInfinityWeight; + _positiveInfinityWeight = positiveInfinityWeight; + _pendingHistoricalBoundaryRepair = false; + invalidateDerivedCaches(); + } + + private NonFiniteAwareTDigest boundaryPreparedView() { + if (!_pendingHistoricalBoundaryRepair) { + return this; + } + NonFiniteAwareTDigest prepared = copy(); + prepared.prepareHistoricalBoundaries(); + return prepared; + } + + private void restoreBoundaryNormalizationAfterFailedMerge() { + if (_negativeInfinityWeight == 0.0 && _positiveInfinityWeight == 0.0) { + _finiteDigest.resetBoundaryNormalization(); + } + } + + @Nullable + PercentileTDigestAccumulator finiteDigestIfNoTails() { + if (_negativeInfinityWeight == 0.0 && _positiveInfinityWeight == 0.0) { + invalidateDerivedCaches(); + return _finiteDigest; + } + return null; + } + + private void mutated() { + _originalFractionalBytes = null; + invalidateDerivedCaches(); + } + + @Override + public void compress() { + _finiteDigest.compress(); + invalidateDerivedCaches(); + } + + @Override + public double getTotalWeight() { + return _finiteDigest.getTotalWeight() + _negativeInfinityWeight + _positiveInfinityWeight; + } + + @Override + double getHistoricalFractionalBoundaryMean(boolean lowerBoundary) { + return _finiteDigest.getHistoricalFractionalBoundaryMean(lowerBoundary); + } + + @Override + double getFreshFractionalBoundaryMean(boolean lowerBoundary) { + return _finiteDigest.getFreshFractionalBoundaryMean(lowerBoundary); + } + + @Override + public boolean hasValidStatistics() { + return _finiteDigest.hasValidStatistics(); + } + + @Override + public double cdf(double value) { + if (Double.isNaN(value) || Double.isInfinite(value)) { + throw new IllegalArgumentException(String.format("Invalid value: %f", value)); + } + double size = getTotalWeight(); + if (size == 0.0 || !hasValidStatistics()) { + return Double.NaN; + } + double finiteSize = _finiteDigest.getTotalWeight(); + double finiteWeight = finiteSize == 0L ? 0.0 : _finiteDigest.cdf(value) * finiteSize; + return (_negativeInfinityWeight + finiteWeight) / size; + } + + @Override + public double quantile(double quantile) { + if (Double.isNaN(quantile) || quantile < 0.0 || quantile > 1.0) { + throw new IllegalArgumentException("q should be in [0,1], got " + quantile); + } + double size = getTotalWeight(); + if (size == 0.0 || !hasValidStatistics()) { + return Double.NaN; + } + double index = quantile * size; + if (_negativeInfinityWeight > 0L && index < _negativeInfinityWeight) { + return Double.NEGATIVE_INFINITY; + } + if (_positiveInfinityWeight > 0L && index >= size - _positiveInfinityWeight) { + return Double.POSITIVE_INFINITY; + } + double finiteSize = _finiteDigest.getTotalWeight(); + if (finiteSize == 0L) { + return _negativeInfinityWeight > 0L ? Double.NEGATIVE_INFINITY : Double.POSITIVE_INFINITY; + } + double finiteQuantile = (index - _negativeInfinityWeight) / finiteSize; + return _finiteDigest.quantile(Math.max(0.0, Math.min(finiteQuantile, 1.0))); + } + + @Override + public Collection<Centroid> centroids() { + return getAllCentroids(); + } + + @Override + public double compression() { + return _finiteDigest.compression(); + } + + @Override + public int maxSerializedByteSize() { + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.length; + } + if (_negativeInfinityWeight == 0.0 && _positiveInfinityWeight == 0.0) { + return _finiteDigest.maxSerializedByteSize(); + } + int finiteByteSize = _finiteDigest.maxSerializedByteSize(); + PercentileTDigestAccumulator finiteDigest = _finiteDigest; + if (finiteDigest.hasOriginalFractionalPayload()) { + // Infinite tails can make retained fractional boundaries interior; their finite view may split endpoints. + finiteByteSize = Math.max(finiteByteSize, + TDigestCodec.getMaxLegacyCompatibleByteSize(compression(), finiteDigest.getCentroidCountUpperBound())); + } + boolean hasFiniteValues = _finiteDigest.getTotalWeight() > 0.0; + int negativeInfinityCount = getInfinityCentroidCount(_negativeInfinityWeight, true, + !hasFiniteValues && _positiveInfinityWeight == 0.0); + int positiveInfinityCount = getInfinityCentroidCount(_positiveInfinityWeight, + !hasFiniteValues && _negativeInfinityWeight == 0.0, true); + int tailCount = Math.addExact(negativeInfinityCount, positiveInfinityCount); + int tailByteSize = Math.multiplyExact(TDigestCodec.VERBOSE_CENTROID_SIZE, tailCount); + long finiteCount = hasFiniteValues ? finiteDigest.getCentroidCountUpperBound() : 0L; + int compatibleBound = + TDigestCodec.getMaxLegacyCompatibleByteSize(compression(), Math.addExact(finiteCount, tailCount)); + return Math.min(Math.addExact(finiteByteSize, tailByteSize), compatibleBound); + } + + @Override + public int centroidCount() { + return getAllCentroids().size(); + } + + @Override + public double getMin() { + if (_pendingHistoricalBoundaryRepair) { + return _historicalMin; + } + if (_negativeInfinityWeight > 0L) { + return Double.NEGATIVE_INFINITY; + } + if (!_finiteDigest.isEmpty()) { + return _finiteDigest.getMin(); + } + return Double.POSITIVE_INFINITY; + } + + @Override + public double getMax() { + if (_pendingHistoricalBoundaryRepair) { + return _historicalMax; + } + if (_positiveInfinityWeight > 0L) { + return Double.POSITIVE_INFINITY; + } + if (!_finiteDigest.isEmpty()) { + return _finiteDigest.getMax(); + } + return Double.NEGATIVE_INFINITY; + } + + private List<Centroid> getFiniteCentroids() { + if (_finiteCentroids == null) { + // Centroid reads use the native working-compression view; only serialization applies public compression. + _finiteCentroids = new ArrayList<>(_finiteDigest.centroids()); + } + return _finiteCentroids; + } + + private List<Centroid> getAllCentroids() { + List<Centroid> finiteCentroids = getFiniteCentroids(); + boolean hasFiniteValues = _finiteDigest.getTotalWeight() > 0.0; + List<Centroid> centroids = new ArrayList<>(finiteCentroids.size() + 6); + appendInfinityCentroids(centroids, Double.NEGATIVE_INFINITY, _negativeInfinityWeight, true, + !hasFiniteValues && _positiveInfinityWeight == 0L); + centroids.addAll(finiteCentroids); + appendInfinityCentroids(centroids, Double.POSITIVE_INFINITY, _positiveInfinityWeight, + !hasFiniteValues && _negativeInfinityWeight == 0L, true); + return centroids; + } + + @Override + public byte[] serialize() { + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.clone(); + } + if (_pendingHistoricalBoundaryRepair) { + return boundaryPreparedView().serialize(); + } + if (_negativeInfinityWeight == 0.0 && _positiveInfinityWeight == 0.0) { + byte[] bytes = _finiteDigest.serialize(); + invalidateDerivedCaches(); + return bytes; + } + // Repair boundaries only after enclosing infinity tails establish the complete distribution's extrema. + _finiteDigest.compress(); + invalidateDerivedCaches(); + int finiteCentroidCount = _finiteDigest.centroidCount(); + boolean hasFiniteValues = _finiteDigest.getTotalWeight() > 0.0; + int negativeInfinityCentroidCount = getInfinityCentroidCount(_negativeInfinityWeight, true, + !hasFiniteValues && _positiveInfinityWeight == 0.0); + int positiveInfinityCentroidCount = getInfinityCentroidCount(_positiveInfinityWeight, + !hasFiniteValues && _negativeInfinityWeight == 0.0, true); + int centroidCount = Math.addExact(finiteCentroidCount, + Math.addExact(negativeInfinityCentroidCount, positiveInfinityCentroidCount)); + double[] means = new double[Math.addExact(centroidCount, 2)]; + double[] weights = new double[means.length]; + _finiteDigest.copyCentroids(means, weights, negativeInfinityCentroidCount); + appendInfinityCentroids(means, weights, 0, Double.NEGATIVE_INFINITY, _negativeInfinityWeight, true, + !hasFiniteValues && _positiveInfinityWeight == 0.0); + appendInfinityCentroids(means, weights, negativeInfinityCentroidCount + finiteCentroidCount, + Double.POSITIVE_INFINITY, _positiveInfinityWeight, + !hasFiniteValues && _negativeInfinityWeight == 0.0, true); + centroidCount = PercentileTDigestAccumulator.normalizeSerializedBoundaries( + means, weights, centroidCount, getMin(), getMax()); + return TDigestCodec.serializeCompatibleCentroids(compression(), getMin(), getMax(), means, weights, centroidCount, + _finiteDigest.hasInheritedFractionalBoundaryEncoding(means, weights, centroidCount)); + } + + private void invalidateDerivedCaches() { + _finiteCentroids = null; + } + + private static void appendInfinityCentroids(List<Centroid> centroids, double value, double weight, + boolean unitWeightAtStart, boolean unitWeightAtEnd) { + appendInfinityCentroids(centroids, null, null, 0, value, weight, unitWeightAtStart, unitWeightAtEnd); + } + + // A null destination counts the same split without allocating centroids or temporary arrays. + private static int appendInfinityCentroids(@Nullable List<Centroid> centroids, @Nullable double[] means, + @Nullable double[] weights, int offset, double value, double weight, boolean unitWeightAtStart, + boolean unitWeightAtEnd) { + if (!(weight > 0.0)) { + return 0; + } + int count = 0; + if (unitWeightAtStart && weight >= (unitWeightAtEnd ? 2.0 : 1.0)) { + appendInfinityCentroid(centroids, means, weights, offset + count++, value, 1.0); + weight -= 1.0; Review Comment: **`weight -= 1.0` is a silent no-op for tail weights >= 2^53.** At weights >= 2^53 the ULP of `weight` is >= 2, so subtracting 1.0 rounds back to the original value. After emitting the unit boundary centroid(s), the residual centroid then still carries the full mass, so the serialized tail contains up to 2 units more mass than `getTotalWeight()` reports. `addInfinityWeight` already guards the symmetric absorption case on addition ("would absorb positive mass") — this decrement path deserves the same guard (or an exactness check) so the split cannot silently create mass. The old long-based bookkeeping with `Math.addExact` was exact here. ########## pinot-segment-local/src/main/java/org/apache/pinot/segment/local/customobject/tdigest/SerializedTDigest.java: ########## @@ -42,6 +41,6 @@ public int compareTo(SerializedTDigest other) { @Override public String toString() { - return BytesUtils.toHexString(CustomSerDeUtils.TDIGEST_SER_DE.serialize(_tDigest)); + return BytesUtils.toHexString(_tDigest.serialize()); Review Comment: **`toString()` can destructively recompress the digest after `compareTo` computed the sort key.** For reducer accumulators, `compareTo` -> `quantile()` flushes at the working (2x) compression and leaves `_publiclyCompressed == false`; `toString()` -> `serialize()` -> `compress()` then runs `recompressCentroids(_compression)`, destructively replacing the centroid set. The hex returned to the client therefore encodes a different (further-compressed) digest than the one that produced the ORDER BY position, and a `compareTo` after rendering can flip sign for near-ties — under tdunning, `quantile` and `asBytes` read the same state. Materiality is low (all trim/sort comparisons complete before rendering in the current query flow), but a side-effecting `toString`/non-idempotent comparison on the same object is fragile; consider serializing from a snapshot or making `quantile()` on a to-be-serialized digest use the public compression. ########## pinot-segment-local/src/main/java/org/apache/pinot/segment/local/customobject/tdigest/PercentileTDigestAccumulator.java: ########## @@ -0,0 +1,1703 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +/* + * Portions of the CDF, quantile, weighted-average and K1 weight-limit implementation are adapted from + * MergingDigest, AbstractTDigest and ScaleFunction in t-digest 3.3 (https://github.com/tdunning/t-digest). + * Modified for Pinot's primitive centroid storage, lazy serialized inputs and legacy boundary handling. + * Original license notice for these portions: + * + * Licensed to Ted Dunning under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.pinot.segment.local.customobject.tdigest; + +import java.nio.ByteBuffer; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Collection; +import java.util.List; +import org.apache.pinot.segment.local.customobject.tdigest.TDigestCodec.SerializedTDigestMetadata; + + + +/// Accumulates raw values and serialized TDigests into primitive centroids using Pinot's K1 weight-limit rule. +/// +/// Raw values are sorted in small batches; serialized centroids retain their existing sorted order. Both are +/// linearly merged with accumulated centroids. K1 preserves Pinot's existing middle-quantile behavior, and the +/// legacy verbose and small encodings preserve stored data and mixed-version wire compatibility. +/// +/// Serialized state keeps its first digest pending and allocates centroid buffers only when another input must be +/// merged or the digest is queried. Raw-value buffers remain unallocated until a raw value is added. +/// +/// Historical corrupted payloads remain opaque and byte-exact, with NaN statistics. Their reported historical mass +/// may be invalid; they cannot be mutated or mixed with another distribution. Such an operation fails before +/// changing aggregate mass, because neither invented statistics nor synthetic NaN bytes are safe for old readers. +/// Unchanged historical fractional boundaries retain their original bytes, including when configured compression +/// differs. After a merge, inherited fractional endpoints retain precise verbose weights and finite statistics when +/// unit endpoints cannot preserve their mass. These keep the historical t-digest 3.3 assertion-enabled reader and +/// recompression limitation. Newly introduced unsupported fractional endpoints are rejected. +/// +/// Instances are externally serialized per aggregation or group key and are not safe for concurrent mutation. +public final class PercentileTDigestAccumulator extends TDigest { + private static final int MIN_RAW_BUFFER_SIZE = 256; + private static final int MAX_RAW_BUFFER_SIZE = 10_000; + private static final int MIN_COMPRESSION = 10; + private static final int TWO_LEVEL_COMPRESSION_MULTIPLIER = 2; + private static final int MIN_INCREMENTAL_COMPRESSION = 50; + + private final double _compression; + private final boolean _useTwoLevelRawCompression; + private final double _workingCompression; + private final double _publicMaxWeightScale; + private final double _workingMaxWeightScale; + private final int _centroidCapacity; + private final int _pendingCentroidCapacity; + private double[] _rawValues; + private double[] _centroidMeans; + private double[] _centroidWeights; + private double[] _outputMeans; + private double[] _outputWeights; + private double[] _incomingMeans; + private double[] _incomingWeights; + private double[] _sortedIncomingMeans; + private double[] _sortedIncomingWeights; + private SerializedTDigestInput _serializedTDigestInput; + private byte[] _pendingSerializedTDigest; + private byte[] _originalFractionalBytes; + private byte[] _originalBoundaryBytes; + private double _historicalFractionalMinMean = Double.NaN; + private double _historicalFractionalMaxMean = Double.NaN; + private double _freshFractionalMinMean = Double.NaN; + private double _freshFractionalMaxMean = Double.NaN; + private SerializedTDigestMetadata _pendingSerializedMetadata; + private boolean _legacyDegraded; + private boolean _hasFractionalWeights; + private boolean _deferBoundaryNormalization; + + private int _numRawValues; + private int _numCentroids; + private int _numIncomingCentroids; + private int _mergeCount; + private boolean _publiclyCompressed; + private boolean _incomingCentroidsSorted = true; + private double _totalWeight; + private double _incomingWeight; + private double _min = Double.POSITIVE_INFINITY; + private double _max = Double.NEGATIVE_INFINITY; + + public PercentileTDigestAccumulator(int compression) { + this((double) compression, true, true); + } + + public PercentileTDigestAccumulator(double compression) { + this(compression, true, true); + } + + private PercentileTDigestAccumulator(double compression, boolean allocateRawBuffer, boolean allocateMergeBuffers) { + this(compression, allocateRawBuffer, allocateMergeBuffers, false); + } + + private PercentileTDigestAccumulator(double compression, boolean allocateRawBuffer, boolean allocateMergeBuffers, + boolean useTwoLevelRawCompression) { + TDigestCodec.validateCompression(compression); + _useTwoLevelRawCompression = useTwoLevelRawCompression; + _compression = Math.max(MIN_COMPRESSION, compression); + _workingCompression = TWO_LEVEL_COMPRESSION_MULTIPLIER * _compression; + _publicMaxWeightScale = calculateK1MaxWeightScale(_compression); + _workingMaxWeightScale = calculateK1MaxWeightScale(_workingCompression); + long roundedCompression = (long) Math.ceil(_compression); + _centroidCapacity = TDigestCodec.getDefaultCentroidCapacity(_compression); + _pendingCentroidCapacity = getPendingCentroidCapacity(_centroidCapacity, useTwoLevelRawCompression); + if (allocateRawBuffer) { + _rawValues = new double[getInitialRawBufferSize(roundedCompression)]; + } + if (allocateMergeBuffers) { + int initialCapacity = Math.min(_centroidCapacity, MIN_RAW_BUFFER_SIZE); + _centroidMeans = new double[initialCapacity]; + _centroidWeights = new double[initialCapacity]; + _outputMeans = new double[initialCapacity]; + _outputWeights = new double[initialCapacity]; + } + } + + /// Creates a lazy digest with the legacy two-level compression used by Star-tree and general digest aggregation. + public static PercentileTDigestAccumulator forLegacyAggregation(double compression) { + return new PercentileTDigestAccumulator(compression, false, false, true); + } + + /// Reads and retains one legacy digest, preserving the general aggregation compression policy. + public static PercentileTDigestAccumulator fromBytes(byte[] bytes) { + return fromBytes(ByteBuffer.wrap(bytes)); + } + + /// Reads one payload and advances the buffer past it. + public static PercentileTDigestAccumulator fromBytes(ByteBuffer bytes) { + SerializedTDigestInput input = new SerializedTDigestInput(); + input.reset(bytes); + return fromBytes(input); + } + + /// Retains an already validated input without rescanning its centroid bytes. + public static PercentileTDigestAccumulator fromBytes(SerializedTDigestInput input) { + PercentileTDigestAccumulator digest = forLegacyAggregation(input.getCompression()); + digest.addSerializedTDigest(input); + return digest; + } + + public static PercentileTDigestAccumulator forReduction(double compression) { + return new PercentileTDigestAccumulator(compression, false, false); + } + + /// Snapshots buffered inputs and centroid state without materializing or compressing the source. Mutable buffers + /// are copied; immutable retained bytes and metadata are shared. Merge and sort scratch buffers remain lazy. + PercentileTDigestAccumulator copy() { + PercentileTDigestAccumulator copy = + new PercentileTDigestAccumulator(_compression, false, false, _useTwoLevelRawCompression); + copy._rawValues = _rawValues != null ? _rawValues.clone() : null; + copy._centroidMeans = _numCentroids > 0 ? Arrays.copyOf(_centroidMeans, _numCentroids) : null; + copy._centroidWeights = _numCentroids > 0 ? Arrays.copyOf(_centroidWeights, _numCentroids) : null; + copy._incomingMeans = _numIncomingCentroids > 0 ? Arrays.copyOf(_incomingMeans, _numIncomingCentroids) : null; + copy._incomingWeights = _numIncomingCentroids > 0 ? Arrays.copyOf(_incomingWeights, _numIncomingCentroids) : null; + copy._pendingSerializedTDigest = _pendingSerializedTDigest; + copy._pendingSerializedMetadata = _pendingSerializedMetadata; + copy._originalFractionalBytes = _originalFractionalBytes; + copy._originalBoundaryBytes = _originalBoundaryBytes; + copy._historicalFractionalMinMean = _historicalFractionalMinMean; + copy._historicalFractionalMaxMean = _historicalFractionalMaxMean; + copy._freshFractionalMinMean = _freshFractionalMinMean; + copy._freshFractionalMaxMean = _freshFractionalMaxMean; + copy._legacyDegraded = _legacyDegraded; + copy._hasFractionalWeights = _hasFractionalWeights; + copy._deferBoundaryNormalization = _deferBoundaryNormalization; + copy._numRawValues = _numRawValues; + copy._numCentroids = _numCentroids; + copy._numIncomingCentroids = _numIncomingCentroids; + copy._mergeCount = _mergeCount; + copy._publiclyCompressed = _publiclyCompressed; + copy._incomingCentroidsSorted = _incomingCentroidsSorted; + copy._totalWeight = _totalWeight; + copy._incomingWeight = _incomingWeight; + copy._min = _min; + copy._max = _max; + return copy; + } + + /// Copies primitive centroids without encoding and decoding them. Zero weights are ignored, and unordered + /// means are sorted when initializing a stored, already compressed distribution; caller arrays are not mutated. + /// Subsequent inputs use the normal buffered merge. Positive weights and consistent extrema are required. + public void addCentroids(double[] means, double[] weights, int count, double min, double max, + boolean alreadyCompressed) { + addCentroids(means, weights, count, min, max, alreadyCompressed, false); + } + + /// Adds validated legacy centroids without marking their existing fractional mass as fresh input. + void addHistoricalCentroids(double[] means, double[] weights, int count, double min, double max, + boolean alreadyCompressed) { + addCentroids(means, weights, count, min, max, alreadyCompressed, true); + } + + private void addCentroids(double[] means, double[] weights, int count, double min, double max, + boolean alreadyCompressed, boolean historical) { + requireMutable(); + if (count < 0 || count > means.length || count > weights.length) { + throw new IllegalArgumentException("Invalid TDigest centroid count: " + count); + } + double totalWeight = 0.0; + int positiveCount = 0; + boolean sorted = true; + boolean fractionalWeights = false; + double previousMean = Double.NEGATIVE_INFINITY; + for (int i = 0; i < count; i++) { + double weight = weights[i]; + if (weight == 0.0) { + continue; + } + double mean = means[i]; + if (!(weight > 0.0) || !Double.isFinite(weight) || !(mean >= min && mean <= max)) { + throw new IllegalArgumentException("Invalid TDigest centroid"); + } + totalWeight += weight; + sorted &= mean >= previousMean; + previousMean = mean; + fractionalWeights |= weight != Math.rint(weight); + positiveCount++; + } + checkTotalWeight(getTotalWeight() + totalWeight); + if (positiveCount == 0) { + return; + } + prepareForMutation(); + _originalFractionalBytes = null; + _hasFractionalWeights |= fractionalWeights; + if (!historical && fractionalWeights) { + for (int i = 0; i < count; i++) { + recordFreshFractionalMean(means[i], weights[i]); + } + } + if (hasNoInputs() && alreadyCompressed) { + ensureCentroidCapacity(Math.addExact(positiveCount, 2)); + for (int i = 0; i < count; i++) { + if (weights[i] > 0.0) { + _centroidMeans[_numCentroids] = means[i]; + _centroidWeights[_numCentroids++] = weights[i]; + } + } + if (!sorted) { + sortCentroids(_centroidMeans, _centroidWeights, _numCentroids); + } + _totalWeight = totalWeight; + _min = min; + _max = max; + if (!historical) { + normalizeBoundaryCentroids(); + } + _publiclyCompressed = true; + } else { + if (sorted && positiveCount == count) { + bufferIncomingCentroids(means, weights, count, totalWeight, min, max); + } else { + // Caller-owned arrays are copied only when filtering or sorting is necessary. + double[] sortedMeans = new double[positiveCount]; + double[] sortedWeights = new double[positiveCount]; + int index = 0; + for (int i = 0; i < count; i++) { + if (weights[i] > 0.0) { + sortedMeans[index] = means[i]; + sortedWeights[index++] = weights[i]; + } + } + if (!sorted) { + sortCentroids(sortedMeans, sortedWeights, positiveCount); + } + bufferIncomingCentroids(sortedMeans, sortedWeights, positiveCount, totalWeight, min, max); + } + _min = Math.min(_min, min); + _max = Math.max(_max, max); + } + } + + public void add(double[] values, int from, int toExclusive) { + if (from >= toExclusive) { + return; + } + requireMutable(); + prepareForMutation(); + while (from < toExclusive) { + ensureRawBuffer(); + if (_numRawValues == _rawValues.length + || _numRawValues + _numIncomingCentroids == getPendingInputLimit()) { + flush(); + } + int numValues = Math.min(toExclusive - from, Math.min(_rawValues.length - _numRawValues, + getPendingInputLimit() - _numRawValues - _numIncomingCentroids)); + int rawOffset = _numRawValues; + for (int i = 0; i < numValues; i++) { + double value = values[from + i]; + if (Double.isNaN(value)) { + throw new IllegalArgumentException("Cannot add NaN to t-digest"); + } + _rawValues[rawOffset + i] = value; + } + _originalFractionalBytes = null; + from += numValues; + _numRawValues += numValues; + if (_numRawValues == _rawValues.length + && (!_useTwoLevelRawCompression || _rawValues.length == _pendingCentroidCapacity) + || _numRawValues + _numIncomingCentroids == getPendingInputLimit()) { + flush(); + } + } + } + + @Override + public void add(double value) { + if (Double.isNaN(value)) { + throw new IllegalArgumentException("Cannot add NaN to t-digest"); + } + requireMutable(); + prepareForMutation(); + ensureRawBuffer(); + if (_numRawValues == _rawValues.length + || _numRawValues + _numIncomingCentroids == getPendingInputLimit()) { + flush(); + } + _originalFractionalBytes = null; + _rawValues[_numRawValues++] = value; + } + + @Override + public void add(double value, double weight) { + if (weight == 1) { + add(value); + return; + } + if (Double.isNaN(value)) { + throw new IllegalArgumentException("Cannot add NaN to t-digest"); + } + if (!(weight > 0.0) || !Double.isFinite(weight)) { + throw new IllegalArgumentException("TDigest weight must be positive: " + weight); + } + requireMutable(); + prepareForMutation(); + bufferIncomingCentroid(value, weight); + _originalFractionalBytes = null; + _hasFractionalWeights |= weight != Math.rint(weight); + recordFreshFractionalMean(value, weight); + _min = Math.min(_min, value); + _max = Math.max(_max, value); + } + + @Override + public void add(TDigest other) { + if (other.hasValidStatistics() && other.getTotalWeight() == 0.0) { + return; + } + requireMutable(); + if (other instanceof PercentileTDigestAccumulator) { + addAccumulator((PercentileTDigestAccumulator) other); + return; + } + if (other instanceof NonFiniteAwareTDigest) { + PercentileTDigestAccumulator finiteDigest = ((NonFiniteAwareTDigest) other).finiteDigestIfNoTails(); + if (finiteDigest != null) { + addAccumulator(finiteDigest); + return; + } + } + if (!other.hasValidStatistics()) { + materializePendingSerializedTDigest(); + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(ByteBuffer.wrap(other.serialize()), false); + addSerializedTDigest(input); + return; + } + double otherWeight = other.getTotalWeight(); + checkTotalWeight(getTotalWeight() + otherWeight); + Collection<Centroid> centroids = other.centroids(); + boolean wrapped = other instanceof NonFiniteAwareTDigest; + int capacity = wrapped ? Math.addExact(centroids.size(), 2) : centroids.size(); + double[] means = new double[capacity]; + double[] weights = new double[capacity]; + int count = 0; + for (Centroid centroid : centroids) { + double weight = centroid.weight(); + if (weight == 0.0) { + continue; + } + if (!(weight > 0.0) || !Double.isFinite(weight) || Double.isNaN(centroid.mean())) { + throw new IllegalArgumentException("Invalid TDigest centroid"); + } + means[count] = centroid.mean(); + weights[count++] = weight; + } + if (wrapped) { + count = normalizeSerializedBoundaries(means, weights, count, other.getMin(), other.getMax()); + } + // Validate all generic centroids before changing the target, then use the same whole-source primitive path. + addCentroids(means, weights, count, other.getMin(), other.getMax(), false, wrapped); + inheritHistoricalFractionalBoundaries(other); + } + + private void addAccumulator(PercentileTDigestAccumulator other) { + if (other == this) { + compress(); + prepareForMutation(); + if (_numCentroids == 0) { + return; + } + // Snapshot only self-merges; fresh fractional mass need not be representable by a legacy wire payload. + addHistoricalCentroids(Arrays.copyOf(_centroidMeans, _numCentroids), + Arrays.copyOf(_centroidWeights, _numCentroids), _numCentroids, _min, _max, false); + return; + } + if (_deferBoundaryNormalization && !other._deferBoundaryNormalization) { + other.prepareForMutation(); + } + if (other._originalBoundaryBytes != null) { + if (getTotalWeight() != 0.0) { + throw unrepresentableBoundaryMutation(); + } + addSerializedTDigest(other._originalBoundaryBytes); + return; + } + if (other._originalFractionalBytes != null && getTotalWeight() == 0.0) { + materializePendingSerializedTDigest(); + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(other._originalFractionalBytes); + input._retainedBytes = other._originalFractionalBytes; + addSerializedTDigest(input); + return; + } + if (other._pendingSerializedTDigest != null) { + // Materialize before resetting the reusable reader: materialization uses that same reader. + materializePendingSerializedTDigest(); + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(other._pendingSerializedTDigest, other._pendingSerializedMetadata); + input._retainedBytes = other._pendingSerializedTDigest; + addSerializedTDigest(input); + return; + } + if (other._legacyDegraded) { + addSerializedTDigest(other.serialize()); + return; + } + + // MergingDigest compressed merge sources in place. Aggregation owns these mutable inputs exclusively. + other.compress(); + other.prepareForMutation(); + prepareForMutation(); + checkTotalWeight(_totalWeight + _incomingWeight + _numRawValues + other._totalWeight); + _hasFractionalWeights |= other._hasFractionalWeights; + inheritHistoricalFractionalBoundaries(other); + bufferIncomingCentroids(other._centroidMeans, other._centroidWeights, other._numCentroids, other._totalWeight, + other._min, other._max); + if (other._totalWeight > 0.0) { + _originalFractionalBytes = null; + } + _min = Math.min(_min, other._min); + _max = Math.max(_max, other._max); + } + + public void addSerializedTDigest(byte[] bytes) { + if (_legacyDegraded || _originalBoundaryBytes != null) { + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(bytes); + addSerializedTDigest(input); + return; + } + requireMutable(); + materializePendingSerializedTDigest(); + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(bytes); + addSerializedTDigest(input); + } + + private SerializedTDigestInput getSerializedTDigestInput() { + if (_serializedTDigestInput == null) { + _serializedTDigestInput = new SerializedTDigestInput(); + } + return _serializedTDigestInput; + } + + public void addSerializedTDigest(SerializedTDigestInput input) { + if (ignoreEmptyInputForDegradedState(input)) { + return; + } + requireMutable(); + if (prepareSerializedTDigest(input)) { + mergeSerializedTDigest(input, false); + } + } + + public void addSerializedTDigestDirect(SerializedTDigestInput input) { + if (ignoreEmptyInputForDegradedState(input)) { + return; + } + requireMutable(); + if (prepareSerializedTDigest(input)) { + mergeSerializedTDigest(input, true); + } + } + + private boolean ignoreEmptyInputForDegradedState(SerializedTDigestInput input) { + if (!_legacyDegraded && _originalBoundaryBytes == null + && (_pendingSerializedMetadata == null || !_pendingSerializedMetadata.weightedBoundaries())) { + return false; + } + input.inspectMetadata(); + return !input._metadata._needsLegacyFallback && input._metadata._totalWeight == 0.0; + } + + private boolean prepareSerializedTDigest(SerializedTDigestInput input) { + if (_pendingSerializedTDigest == null && hasNoInputs()) { + input.inspectMetadata(); + _hasFractionalWeights |= input._metadata._fractionalWeights; + _pendingSerializedTDigest = input.retainBytes(); + _pendingSerializedMetadata = input.getMetadata(); + if (input.recordHistoricalFractionalBoundaries(this)) { + // Alias the retained snapshot; historical fractional bytes remain byte-exact after read-only materialization. + _originalFractionalBytes = _pendingSerializedTDigest; + } + if (input._metadata._needsLegacyFallback) { + _legacyDegraded = true; + } + return false; + } + materializePendingSerializedTDigest(); + input.decode(); + if (input._metadata._needsLegacyFallback) { + if (hasNoInputs()) { + _pendingSerializedTDigest = input.retainBytes(); + _pendingSerializedMetadata = input.getMetadata(); + _legacyDegraded = true; + return false; + } + throw corruptedMutation(); + } + if (!_deferBoundaryNormalization) { + input.normalizeBoundaries(); + } + prepareForMutation(); + checkTotalWeight(_totalWeight + _incomingWeight + _numRawValues + input._metadata._totalWeight); + input.recordHistoricalFractionalBoundaries(this); + if (input._metadata._totalWeight > 0.0) { + _originalFractionalBytes = null; + } + _hasFractionalWeights |= input._metadata._fractionalWeights; + return true; + } + + private boolean hasNoInputs() { + return _numCentroids == 0 && _totalWeight == 0.0 && _numRawValues == 0 && _numIncomingCentroids == 0; + } + + private void mergeSerializedTDigest(SerializedTDigestInput input, boolean direct) { + if (input._metadata._totalWeight == 0.0) { + return; + } + if (hasNoInputs()) { + if (input._metadata._totalWeight != 0.0) { + ensureCentroidCapacity(input._numCentroids); + System.arraycopy(input._means, 0, _centroidMeans, 0, input._numCentroids); + System.arraycopy(input._weights, 0, _centroidWeights, 0, input._numCentroids); + _numCentroids = input._numCentroids; + _totalWeight = input._metadata._totalWeight; + _min = input._metadata._min; + _max = input._metadata._max; + // Stored centroids already went through the writer's public compression. Reading or repairing their + // endpoints needs another K1 pass only when the output compression differs from the stored compression. + _publiclyCompressed = _compression == input._metadata._compression; + } + return; + } + if (direct) { + flush(); + boolean runBackwards = (_mergeCount++ & 1) != 0; + if (!mergeSorted(input._means, input._weights, input._numCentroids, input._metadata._totalWeight, + getIncrementalCompression(), runBackwards)) { + return; + } + } else { + bufferIncomingCentroids(input._means, input._weights, input._numCentroids, input._metadata._totalWeight, + input._metadata._min, input._metadata._max); + } + _min = Math.min(_min, input._metadata._min); + _max = Math.max(_max, input._metadata._max); + } + + @Override + public void compress() { + if (_legacyDegraded) { + return; + } + materializePendingSerializedTDigest(); + if (_useTwoLevelRawCompression) { + // Public compression and serialization both retain the legacy working pass before final compression. + flush(); + } + if (_numRawValues > 0 || _numIncomingCentroids > 0) { + // SQL reduction combines pending values and existing centroids directly at public compression, matching + // MergingDigest 3.3. Its writer avoids an extra working-compression pass that can increase reducer rank error. + flush(_compression); + _publiclyCompressed = true; + } else if (!_publiclyCompressed && _numCentroids > 0) { + recompressCentroids(_compression); + _publiclyCompressed = true; + } + } + + @Override + public boolean hasValidStatistics() { + return !_legacyDegraded; + } + + @Override + public double getTotalWeight() { + return _pendingSerializedTDigest != null ? _pendingSerializedMetadata.totalWeight() + : _totalWeight + _incomingWeight + _numRawValues; + } + + /// Bounds serialized output centroids from buffered inputs, including up to two boundary splits before a merge + /// and two more before writing, without flushing or using the potentially huge input mass. + public long getCentroidCountUpperBound() { + return _pendingSerializedTDigest != null ? (long) _pendingSerializedMetadata.centroidCount() + 4L + : (long) _numCentroids + _numRawValues + _numIncomingCentroids + 4L; + } + + /// Whether unchanged historical fractional boundaries still require their original byte representation. + public boolean hasOriginalFractionalPayload() { + return _originalFractionalBytes != null; + } + + @Override + public double cdf(double value) { + if (Double.isNaN(value) || Double.isInfinite(value)) { + throw new IllegalArgumentException(String.format("Invalid value: %f", value)); + } + if (_legacyDegraded) { + return Double.NaN; + } + materializePendingSerializedTDigest(); + flush(); + if (_numCentroids == 0) { + return Double.NaN; + } + if (_numCentroids == 1) { + double width = _max - _min; + if (value < _min) { + return 0.0; + } else if (value > _max) { + return 1.0; + } else if (value - _min <= width) { + return 0.5; + } else { + return (value - _min) / width; + } + } + if (value < _min) { + return 0.0; + } + if (value > _max) { + return 1.0; + } + if (value < _centroidMeans[0]) { + double width = _centroidMeans[0] - _min; + if (width > 0.0) { + if (_centroidWeights[0] < 1.0) { + return (_centroidWeights[0] / 2.0 * (value - _min) / width) / _totalWeight; + } + return value == _min ? 0.5 / _totalWeight + : (1.0 + (value - _min) / width * (_centroidWeights[0] / 2.0 - 1.0)) / _totalWeight; + } + return 0.0; + } + int lastIndex = _numCentroids - 1; + if (value > _centroidMeans[lastIndex]) { + double width = _max - _centroidMeans[lastIndex]; + if (width > 0.0) { + if (_centroidWeights[lastIndex] < 1.0) { + return 1.0 - (_centroidWeights[lastIndex] / 2.0 * (_max - value) / width) / _totalWeight; + } + return value == _max ? 1.0 - 0.5 / _totalWeight + : 1.0 - (1.0 + (_max - value) / width * (_centroidWeights[lastIndex] / 2.0 - 1.0)) + / _totalWeight; + } + return 1.0; + } + + double weightSoFar = 0.0; + for (int i = 0; i < lastIndex; i++) { + if (_centroidMeans[i] == value) { + double equalWeight = 0.0; + int equalIndex = i; + while (equalIndex < _numCentroids && _centroidMeans[equalIndex] == value) { + equalWeight += _centroidWeights[equalIndex]; + equalIndex++; + } + return (weightSoFar + equalWeight / 2.0) / _totalWeight; + } + if (_centroidMeans[i] <= value && _centroidMeans[i + 1] > value) { + if (!Double.isFinite(_centroidMeans[i]) || !Double.isFinite(_centroidMeans[i + 1])) { + if (_centroidMeans[i] == Double.NEGATIVE_INFINITY) { + return (weightSoFar + _centroidWeights[i]) / _totalWeight; + } + if (_centroidWeights[i] == 1.0 && _centroidWeights[i + 1] == 1.0) { + return (weightSoFar + 1.0) / _totalWeight; + } + double leftExcludedWeight = _centroidWeights[i] == 1.0 ? 0.5 : 0.0; + double rightExcludedWeight = _centroidWeights[i + 1] == 1.0 ? 0.5 : 0.0; + double interpolatedWeight = (_centroidWeights[i] + _centroidWeights[i + 1]) / 2.0 + - leftExcludedWeight - rightExcludedWeight; + double base = weightSoFar + _centroidWeights[i] / 2.0 + leftExcludedWeight; + return (base + interpolatedWeight * (value - _centroidMeans[i]) + / (_centroidMeans[i + 1] - _centroidMeans[i])) / _totalWeight; + } + double width = _centroidMeans[i + 1] - _centroidMeans[i]; + if (width > 0.0) { + double leftExcludedWeight = getPointMassHalfWeight(i); + double rightExcludedWeight = getPointMassHalfWeight(i + 1); + double halfWeight = (_centroidWeights[i] + _centroidWeights[i + 1]) / 2.0; + double interpolatedWeight = halfWeight - leftExcludedWeight - rightExcludedWeight; + double base = weightSoFar + _centroidWeights[i] / 2.0 + leftExcludedWeight; + return (base + interpolatedWeight * (value - _centroidMeans[i]) / width) / _totalWeight; + } + return (weightSoFar + (_centroidWeights[i] + _centroidWeights[i + 1]) / 2.0) / _totalWeight; + } + weightSoFar += _centroidWeights[i]; + } + if (value == _centroidMeans[lastIndex]) { + return 1.0 - Math.min(1.0, _centroidWeights[lastIndex]) / 2.0 / _totalWeight; + } + throw new IllegalStateException("Unable to compute TDigest CDF"); + } + + @Override + public double quantile(double quantile) { + if (Double.isNaN(quantile) || quantile < 0.0 || quantile > 1.0) { + throw new IllegalArgumentException("q should be in [0,1], got " + quantile); + } + if (_legacyDegraded) { + return Double.NaN; + } + materializePendingSerializedTDigest(); + flush(); + if (_numCentroids == 0) { + return Double.NaN; + } + if (_numCentroids == 1) { + return _centroidMeans[0]; + } + if (quantile == 0.0) { + return _min; + } + if (quantile == 1.0) { + return _max; + } + + double index = quantile * _totalWeight; + double firstWeight = _centroidWeights[0]; + if (firstWeight < 1.0 && index < firstWeight / 2.0) { + return weightedAverage(_min, firstWeight / 2.0 - index, _centroidMeans[0], index); + } + if (firstWeight >= 1.0 && index < 1.0) { + return _min; + } + if (_centroidWeights[0] > 1.0 && index < _centroidWeights[0] / 2.0) { + return _min + (index - 1.0) / (_centroidWeights[0] / 2.0 - 1.0) * (_centroidMeans[0] - _min); + } + int lastIndex = _numCentroids - 1; + double lastWeight = _centroidWeights[lastIndex]; + if (lastWeight < 1.0 && index >= _totalWeight - lastWeight / 2.0) { + return weightedAverage(_centroidMeans[lastIndex], _totalWeight - index, _max, + index - (_totalWeight - lastWeight / 2.0)); + } + if (lastWeight >= 1.0 && index > _totalWeight - 1.0) { + return _max; + } + if (_centroidWeights[lastIndex] > 1.0 + && _totalWeight - index <= _centroidWeights[lastIndex] / 2.0) { + return _max - (_totalWeight - index - 1.0) / (_centroidWeights[lastIndex] / 2.0 - 1.0) + * (_max - _centroidMeans[lastIndex]); + } + + double weightSoFar = _centroidWeights[0] / 2.0; + for (int i = 0; i < lastIndex; i++) { + double halfWeight = (_centroidWeights[i] + _centroidWeights[i + 1]) / 2.0; + if (weightSoFar + halfWeight > index) { + double leftExcludedWeight = getPointMassHalfWeight(i); + if (index - weightSoFar < leftExcludedWeight) { + return _centroidMeans[i]; + } + double rightExcludedWeight = getPointMassHalfWeight(i + 1); + if (weightSoFar + halfWeight - index <= rightExcludedWeight) { + return _centroidMeans[i + 1]; + } + double leftWeight = index - weightSoFar - leftExcludedWeight; + double rightWeight = weightSoFar + halfWeight - index - rightExcludedWeight; + return weightedAverage(_centroidMeans[i], rightWeight, _centroidMeans[i + 1], leftWeight); + } + weightSoFar += halfWeight; + } + double weightToMax = index - _totalWeight - _centroidWeights[lastIndex] / 2.0; + double weightFromLastCentroid = _centroidWeights[lastIndex] / 2.0 - weightToMax; + return weightedAverage(_centroidMeans[lastIndex], weightToMax, _max, weightFromLastCentroid); + } + + private double getPointMassHalfWeight(int index) { + // Adjacent equal means establish a repeated-value plateau. Keep its complete mass out of interpolation at + // either edge; an isolated weighted centroid still represents an unknown distribution around its mean. + double mean = _centroidMeans[index]; + boolean pointMass = _centroidWeights[index] == 1.0 || index > 0 && _centroidMeans[index - 1] == mean + || index + 1 < _numCentroids && _centroidMeans[index + 1] == mean; + return pointMass ? _centroidWeights[index] / 2.0 : 0.0; + } + + private static double weightedAverage(double firstValue, double firstWeight, double secondValue, + double secondWeight) { + if (firstValue > secondValue) { + return weightedAverage(secondValue, secondWeight, firstValue, firstWeight); + } + if (firstWeight == 0.0) { + return secondValue; + } + if (secondWeight == 0.0 || firstValue == secondValue) { + return firstValue; + } + if (!Double.isFinite(firstValue)) { + return firstValue; + } + if (!Double.isFinite(secondValue)) { + return secondValue; + } + double average = (firstValue * firstWeight + secondValue * secondWeight) / (firstWeight + secondWeight); + return Math.max(firstValue, Math.min(average, secondValue)); + } + + @Override + public Collection<Centroid> centroids() { + if (_legacyDegraded) { + byte[] bytes = serialize(); + ByteBuffer encoded = ByteBuffer.wrap(bytes); + SerializedTDigestMetadata metadata = TDigestCodec.readSerializedHeader(encoded); + encoded.position(metadata.centroidOffset()); + List<Centroid> centroids = new ArrayList<>(metadata.centroidCount()); + for (int i = 0; i < metadata.centroidCount(); i++) { + double weight = metadata.centroidSize() == TDigestCodec.VERBOSE_CENTROID_SIZE + ? encoded.getDouble() : encoded.getFloat(); + double mean = metadata.centroidSize() == TDigestCodec.VERBOSE_CENTROID_SIZE + ? encoded.getDouble() : encoded.getFloat(); + centroids.add(new Centroid(mean, weight)); + } + return centroids; + } + materializePendingSerializedTDigest(); + flush(); + List<Centroid> centroids = new ArrayList<>(_numCentroids); + for (int i = 0; i < _numCentroids; i++) { + centroids.add(new Centroid(_centroidMeans[i], _centroidWeights[i])); + } + return centroids; + } + + /// Copies a valid distribution's primitive centroid arrays after flushing buffered inputs. This does not apply + /// public compression or serialize a finite sub-distribution; callers can include surrounding infinity mass. + public void copyCentroids(double[] means, double[] weights, int offset) { + if (_legacyDegraded) { + throw corruptedMutation(); + } + materializePendingSerializedTDigest(); + flush(); + if (means == weights || offset < 0 || offset > means.length - _numCentroids + || offset > weights.length - _numCentroids) { + throw new IllegalArgumentException("Insufficient distinct TDigest centroid arrays"); + } + if (_numCentroids == 0) { + return; + } + System.arraycopy(_centroidMeans, 0, means, offset, _numCentroids); + System.arraycopy(_centroidWeights, 0, weights, offset, _numCentroids); + } + + @Override + public double compression() { + return _compression; + } + + /// Bounds compatible output without flushing. Unchanged historical bytes and retained pending payloads can + /// exceed the newly encoded legacy-capacity bound and retain their actual length. + @Override + public int maxSerializedByteSize() { + if (_legacyDegraded) { + return _pendingSerializedTDigest.length; + } + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.length; + } + if (_originalBoundaryBytes != null) { + return _originalBoundaryBytes.length; + } + long centroidCount = getCentroidCountUpperBound(); + if (!_hasFractionalWeights) { + centroidCount = (long) Math.min(centroidCount, Math.ceil(getTotalWeight())); + } + int maximumBytes = TDigestCodec.getMaxLegacyCompatibleByteSize(_compression, centroidCount); + return _pendingSerializedTDigest != null ? Math.max(maximumBytes, _pendingSerializedTDigest.length) : maximumBytes; + } + + /// Returns the centroid view's count, including repaired endpoints and excluding zero-mass serialized entries. + @Override + public int centroidCount() { + if (_pendingSerializedTDigest != null) { + if (_legacyDegraded) { + return _pendingSerializedMetadata.centroidCount(); + } + if (_pendingSerializedMetadata.isCanonicalPayload()) { + return _pendingSerializedMetadata.centroidCount(); + } + materializePendingSerializedTDigest(); + } + flush(); + return _numCentroids; + } + + @Override + public double getMin() { + if (_pendingSerializedTDigest != null) { + return !_legacyDegraded && _pendingSerializedMetadata.totalWeight() == 0.0 + ? Double.POSITIVE_INFINITY : _pendingSerializedMetadata.min(); + } + flush(); + return _min; + } + + @Override + public double getMax() { + if (_pendingSerializedTDigest != null) { + return !_legacyDegraded && _pendingSerializedMetadata.totalWeight() == 0.0 + ? Double.NEGATIVE_INFINITY : _pendingSerializedMetadata.max(); + } + flush(); + return _max; + } + + @Override + public byte[] serialize() { + if (_legacyDegraded) { + return _pendingSerializedTDigest.clone(); + } + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.clone(); + } + if (_originalBoundaryBytes != null) { + return _originalBoundaryBytes.clone(); + } + if (_pendingSerializedTDigest != null) { + if (!_pendingSerializedMetadata.isCanonicalPayload()) { + materializePendingSerializedTDigest(); + } else { + if (ByteBuffer.wrap(_pendingSerializedTDigest).getInt() == TDigestCodec.VERBOSE_ENCODING) { + byte[] serialized = TDigestCodec.makeLegacyCompatible(_pendingSerializedTDigest, _pendingSerializedMetadata); + return serialized == _pendingSerializedTDigest ? serialized.clone() : serialized; + } + return _pendingSerializedTDigest.clone(); + } + } + if (_originalBoundaryBytes != null) { + return _originalBoundaryBytes.clone(); + } + compress(); + normalizeBoundaryCentroids(); + checkFractionalBoundaryEncoding(); + return TDigestCodec.serializeCompatibleCentroids(_compression, _min, _max, _centroidMeans, _centroidWeights, + _numCentroids, hasInheritedFractionalBoundaryEncoding()); + } + + private void recompressCentroids(double compression) { + normalizeBoundaryCentroids(); + double[] means = _centroidMeans; + double[] weights = _centroidWeights; + int numCentroids = _numCentroids; + double totalWeight = _totalWeight; + _numCentroids = 0; + _totalWeight = 0.0; + boolean runBackwards = (_mergeCount++ & 1) != 0; + if (!mergeSorted(means, weights, numCentroids, totalWeight, compression, runBackwards)) { + throw new IllegalStateException("Cannot recompress an empty TDigest"); + } + } + + private void checkFractionalBoundaryEncoding() { + if (_numCentroids > 0 && !hasInheritedFractionalBoundaryEncoding()) { + TDigestCodec.checkLegacyBoundaryWeights(_numCentroids, _centroidWeights[0], _centroidWeights[_numCentroids - 1]); + } + } + + private boolean hasInheritedFractionalBoundaryEncoding() { + return hasInheritedFractionalBoundaryEncoding(_centroidMeans, _centroidWeights, _numCentroids); + } + + /// Applies the canonical inherited-endpoint check to a complete digest's sorted centroid view. Enclosing writers + /// use this after adding their infinity tails; unsupported fresh global endpoints never inherit permission. + boolean hasInheritedFractionalBoundaryEncoding(double[] means, double[] weights, int count) { + if (count < 0 || count > 0 && (count > means.length || count > weights.length)) { + throw new IllegalArgumentException("Invalid TDigest centroid count: " + count); + } + if (count == 0) { + return false; + } + return TDigestCodec.hasInheritedFractionalBoundaryEncoding(this, count, means[0], weights[0], + means[count - 1], weights[count - 1]); + } + + @Override + double getHistoricalFractionalBoundaryMean(boolean lowerBoundary) { + return lowerBoundary ? _historicalFractionalMinMean : _historicalFractionalMaxMean; + } + + @Override + double getFreshFractionalBoundaryMean(boolean lowerBoundary) { + return lowerBoundary ? _freshFractionalMinMean : _freshFractionalMaxMean; + } + + void recordFreshFractionalMean(double mean, double weight) { + if (weight != Math.rint(weight)) { + inheritFreshFractionalMeans(mean, mean); + } + } + + private void inheritFreshFractionalMeans(double minMean, double maxMean) { + if (!Double.isNaN(minMean)) { + _freshFractionalMinMean = Double.isNaN(_freshFractionalMinMean) ? minMean + : Math.min(_freshFractionalMinMean, minMean); + } + if (!Double.isNaN(maxMean)) { + _freshFractionalMaxMean = Double.isNaN(_freshFractionalMaxMean) ? maxMean + : Math.max(_freshFractionalMaxMean, maxMean); + } + } + + /// Propagates provenance from a validated source without exposing arbitrary endpoint setters. + void inheritHistoricalFractionalBoundaries(TDigest source) { + requireMutable(); + inheritFreshFractionalMeans(source.getFreshFractionalBoundaryMean(true), + source.getFreshFractionalBoundaryMean(false)); + inheritHistoricalFractionalBoundaries(source.getHistoricalFractionalBoundaryMean(true), + source.getHistoricalFractionalBoundaryMean(false)); + } + + void inheritHistoricalFractionalBoundaries(double minMean, double maxMean) { + if (!Double.isNaN(minMean)) { + _historicalFractionalMinMean = Double.isNaN(_historicalFractionalMinMean) ? minMean + : Math.min(_historicalFractionalMinMean, minMean); + } + if (!Double.isNaN(maxMean)) { + _historicalFractionalMaxMean = Double.isNaN(_historicalFractionalMaxMean) ? maxMean + : Math.max(_historicalFractionalMaxMean, maxMean); + } + } + + private void flush() { + // SQL accumulators merge raw inputs incrementally at configured compression. General digest aggregation keeps + // the legacy two-level raw merge, preserving sparse Star-tree query accuracy. Incremental buffer exhaustion or + // centroid/statistics reads use working compression for reducer inputs; final serialization bypasses that pass. + flush(_numIncomingCentroids == 0 && !_useTwoLevelRawCompression + ? getIncrementalCompression() + : _workingCompression); + } + + private double getIncrementalCompression() { + // Very low compression leaves too little centroid headroom for incremental public-compression merges to retain + // t-digest 3.3's accuracy. Keep the two-level strategy for those uncommon configurations. + return _compression >= MIN_INCREMENTAL_COMPRESSION ? _compression : _workingCompression; + } + + private void flush(double compression) { + if (_numIncomingCentroids > 0) { + flushIncoming(compression); + return; + } + if (_numRawValues == 0) { + return; + } + + Arrays.sort(_rawValues, 0, _numRawValues); + double rawMin = _rawValues[0]; + double rawMax = _rawValues[_numRawValues - 1]; + boolean runBackwards = (_mergeCount++ & 1) != 0; + mergeSorted(_rawValues, null, _numRawValues, _numRawValues, compression, runBackwards); + _numRawValues = 0; + _min = Math.min(_min, rawMin); + _max = Math.max(_max, rawMax); + } + + private void flushIncoming(double compression) { + // Match mergeSorted's association before rearranging raw or paired centroid buffers. + checkTotalWeight(_totalWeight + (_incomingWeight + _numRawValues)); + double rawMin = Double.POSITIVE_INFINITY; + double rawMax = Double.NEGATIVE_INFINITY; + if (_numRawValues > 0) { + Arrays.sort(_rawValues, 0, _numRawValues); + rawMin = _rawValues[0]; + rawMax = _rawValues[_numRawValues - 1]; + ensureIncomingCapacity(Math.addExact(_numIncomingCentroids, _numRawValues)); + System.arraycopy(_incomingMeans, 0, _incomingMeans, _numRawValues, _numIncomingCentroids); + System.arraycopy(_incomingWeights, 0, _incomingWeights, _numRawValues, _numIncomingCentroids); + for (int i = 0; i < _numRawValues; i++) { + _incomingMeans[i] = _rawValues[i]; + _incomingWeights[i] = 1.0; + } + _numIncomingCentroids += _numRawValues; + _incomingWeight += _numRawValues; + _numRawValues = 0; + _incomingCentroidsSorted = false; + } + + int incomingCount = _numIncomingCentroids; + double incomingWeight = _incomingWeight; + double[] incomingMeans = _incomingMeans; + double[] incomingWeights = _incomingWeights; + if (!_incomingCentroidsSorted) { + ensureSortedIncomingCapacity(incomingCount); + stableSortIncoming(incomingCount); + incomingMeans = _incomingMeans; + incomingWeights = _incomingWeights; + _incomingCentroidsSorted = true; + } + boolean runBackwards = (_mergeCount & 1) != 0; + if (!mergeSorted(incomingMeans, incomingWeights, incomingCount, incomingWeight, compression, runBackwards)) { + throw new IllegalStateException("Cannot flush an empty TDigest input buffer"); + } + _mergeCount++; + _numIncomingCentroids = 0; + _incomingWeight = 0.0; + _incomingCentroidsSorted = true; + _min = Math.min(_min, rawMin); + _max = Math.max(_max, rawMax); + } + + private boolean mergeSorted(double[] incomingMeans, double[] incomingWeights, int incomingCount, + double incomingWeight, double compression, boolean runBackwards) { + if (incomingWeight == 0.0) { + return false; + } + + double newTotalWeight = _totalWeight + incomingWeight; + checkTotalWeight(newTotalWeight); + normalizeBoundaryCentroids(); + int preferredOutputCapacity = getPreferredOutputCapacity(incomingCount); + ensureOutputCapacity(1); + double totalWeightNormalizer = 1.0 / newTotalWeight; + double weightNormalizer = totalWeightNormalizer / getK1MaxWeightScale(compression); + int incomingIndex = runBackwards ? incomingCount - 1 : 0; + int centroidIndex = runBackwards ? _numCentroids - 1 : 0; + int incomingLimit = runBackwards ? -1 : incomingCount; + int centroidLimit = runBackwards ? -1 : _numCentroids; + int direction = runBackwards ? -1 : 1; + int numInputs = incomingCount + _numCentroids; + int inputIndex = 0; + int numOutputCentroids = 0; + double weightSoFar = 0.0; + double firstInputMean = _numCentroids == 0 ? incomingMeans[0] + : Math.min(incomingMeans[0], _centroidMeans[0]); + double lastInputMean = _numCentroids == 0 ? incomingMeans[incomingCount - 1] + : Math.max(incomingMeans[incomingCount - 1], _centroidMeans[_numCentroids - 1]); + boolean allFiniteMeans = Double.isFinite(firstInputMean) && Double.isFinite(lastInputMean); + boolean sameSignMeans = firstInputMean > 0.0 || lastInputMean < 0.0; + while (incomingIndex != incomingLimit || centroidIndex != centroidLimit) { + double mean; + double weight; + boolean takeIncoming = centroidIndex == centroidLimit || (incomingIndex != incomingLimit + && (runBackwards ? incomingMeans[incomingIndex] > _centroidMeans[centroidIndex] + : incomingMeans[incomingIndex] <= _centroidMeans[centroidIndex])); + if (takeIncoming) { + mean = incomingMeans[incomingIndex]; + weight = incomingWeights == null ? 1.0 : incomingWeights[incomingIndex]; + incomingIndex += direction; + } else { + mean = _centroidMeans[centroidIndex]; + weight = _centroidWeights[centroidIndex]; + centroidIndex += direction; + } + if (numOutputCentroids == 0) { + _outputMeans[0] = mean; + _outputWeights[0] = weight; + numOutputCentroids = 1; + } else { + int currentIndex = numOutputCentroids - 1; + double proposedWeight = _outputWeights[currentIndex] + weight; + double q0 = weightSoFar * totalWeightNormalizer; + double q2 = (weightSoFar + proposedWeight) * totalWeightNormalizer; + double normalizedWeight = proposedWeight * weightNormalizer; + double normalizedWeightSquared = normalizedWeight * normalizedWeight; + boolean canMerge = (allFiniteMeans || canMergeMeans(_outputMeans[currentIndex], mean)) + && normalizedWeightSquared <= q0 * (1.0 - q0) + && normalizedWeightSquared <= q2 * (1.0 - q2); + if (inputIndex == 1 || inputIndex == numInputs - 1) { + canMerge = false; + } + if (canMerge) { + _outputWeights[currentIndex] = proposedWeight; + _outputMeans[currentIndex] = + TDigestCodec.weightedMean(_outputMeans[currentIndex], proposedWeight - weight, mean, weight, + proposedWeight, sameSignMeans); + } else { + weightSoFar += _outputWeights[currentIndex]; + ensureOutputCapacity(Math.max(preferredOutputCapacity, numOutputCentroids + 1)); + _outputMeans[numOutputCentroids] = mean; + _outputWeights[numOutputCentroids] = weight; + numOutputCentroids++; + } + } + inputIndex++; + } + if (runBackwards) { + reverse(_outputMeans, numOutputCentroids); + reverse(_outputWeights, numOutputCentroids); + } + finishMerge(numOutputCentroids, newTotalWeight, compression); + return true; + } + + private void finishMerge(int numOutputCentroids, double totalWeight, double compression) { + double[] temporary = _centroidMeans; + _centroidMeans = _outputMeans; + _outputMeans = temporary; + temporary = _centroidWeights; + _centroidWeights = _outputWeights; + _outputWeights = temporary; + _numCentroids = numOutputCentroids; + _totalWeight = totalWeight; + _publiclyCompressed = compression == _compression; + } + + private void ensureIncomingCapacity(int capacity) { + if (_incomingMeans == null || capacity > _incomingMeans.length) { + int newCapacity = _incomingMeans == null + ? Math.max(capacity, Math.min(_centroidCapacity, _pendingCentroidCapacity)) + : Math.max(capacity, (int) Math.min(_pendingCentroidCapacity, + (long) _incomingMeans.length * TDigestCodec.DEFAULT_MERGE_BUFFER_MULTIPLIER)); + _incomingMeans = _incomingMeans == null ? new double[newCapacity] : Arrays.copyOf(_incomingMeans, newCapacity); + _incomingWeights = + _incomingWeights == null ? new double[newCapacity] : Arrays.copyOf(_incomingWeights, newCapacity); + } + } + + private void stableSortIncoming(int count) { Review Comment: **Two stable-sort implementations for the same paired means/weights data in one class.** `stableSortIncoming` hand-rolls a ~35-line bottom-up merge sort over the parallel arrays, while the decode path in this same file already performs the identical stable paired sort via `it.unimi.dsi.fastutil.Arrays.mergeSort(0, count, comparator, swapper)` (the established Pinot idiom, see `RangeIndexCreator`, `OffHeapSingleTreeBuilder`). A future fix to ordering semantics (NaN, signed zeros) applied to one path will silently miss the other. If the scratch-buffer sort is kept for the flush hot path, route the decode-side sort through it too so exactly one implementation exists. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
