xiangfu0 commented on code in PR #19795: URL: https://github.com/apache/pinot/pull/19795#discussion_r4228358550
########## pinot-segment-local/src/main/java/org/apache/pinot/segment/local/customobject/PercentileTDigestAccumulator.java: ########## @@ -0,0 +1,1662 @@ +/** + * 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; + +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.utils.TDigestUtils; +import org.apache.pinot.segment.local.utils.TDigestUtils.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. +/// +/// 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 SerializedTDigestMetadata _pendingSerializedMetadata; + private boolean _legacyDegraded; + private byte[] _serializedBytesForWrite; + private boolean _hasFractionalWeights; + + private int _numRawValues; + private int _numCentroids; + private int _numIncomingCentroids; + private int _serializedMainCapacity; + private int _serializedBufferCapacity; + 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(positiveCompression(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) { + TDigestUtils.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 = TDigestUtils.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); + } + + public static PercentileTDigestAccumulator forSerializedTDigest(byte[] bytes) { + return new PercentileTDigestAccumulator(readCompression(bytes), false, false); + } + + public static PercentileTDigestAccumulator forSerializedTDigest(SerializedTDigestInput input) { + return new PercentileTDigestAccumulator(input._metadata.compression(), false, false); + } + + public static PercentileTDigestAccumulator forSerializedTDigestWithMergeBuffers(byte[] bytes) { + return new PercentileTDigestAccumulator(readCompression(bytes), false, true); + } + + public static PercentileTDigestAccumulator forSerializedTDigest(ByteBuffer input) { + SerializedTDigestInput serialized = new SerializedTDigestInput(); + serialized.reset(input); + PercentileTDigestAccumulator accumulator = forSerializedTDigest(serialized); + accumulator.addSerializedTDigest(serialized); + return accumulator; + } + + public static PercentileTDigestAccumulator forReduction(double compression) { + return new PercentileTDigestAccumulator(positiveCompression(compression), false, false); + } + + /// 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) { + 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; + } + materializePendingSerializedTDigest(); + _originalFractionalBytes = null; + _hasFractionalWeights |= fractionalWeights; + 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) { + it.unimi.dsi.fastutil.Arrays.mergeSort(0, _numCentroids, + (first, second) -> Double.compare(_centroidMeans[first], _centroidMeans[second]), + (first, second) -> { + double mean = _centroidMeans[first]; + _centroidMeans[first] = _centroidMeans[second]; + _centroidMeans[second] = mean; + double weight = _centroidWeights[first]; + _centroidWeights[first] = _centroidWeights[second]; + _centroidWeights[second] = weight; + }); + } + _totalWeight = totalWeight; + _min = min; + _max = max; + normalizeBoundaryCentroids(); + _publiclyCompressed = true; + } else { + for (int i = 0; i < count; i++) { + if (weights[i] > 0.0) { + bufferIncomingCentroid(means[i], weights[i]); + } + } + _min = Math.min(_min, min); + _max = Math.max(_max, max); + } + } + + public void add(double[] values, int from, int toExclusive) { + if (from >= toExclusive) { + return; + } + requireMutable(); + materializePendingSerializedTDigest(); + 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(); + materializePendingSerializedTDigest(); + 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(); + materializePendingSerializedTDigest(); + bufferIncomingCentroid(value, weight); + _originalFractionalBytes = null; + _hasFractionalWeights |= weight != Math.rint(weight); + _min = Math.min(_min, value); + _max = Math.max(_max, value); + } + + @Override + public void add(TDigest other) { + requireMutable(); + if (other instanceof PercentileTDigestAccumulator) { + addAccumulator((PercentileTDigestAccumulator) other); + return; + } + if (!other.hasValidStatistics()) { + materializePendingSerializedTDigest(); + SerializedTDigestInput input = getSerializedTDigestInput(); + input.reset(ByteBuffer.wrap(TDigestUtils.serialize(other)), false); + addSerializedTDigest(input); + return; + } + materializePendingSerializedTDigest(); + double otherWeight = other.getTotalWeight(); + checkTotalWeight(_totalWeight + _incomingWeight + _numRawValues + otherWeight); + for (Centroid centroid : other.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"); + } + _hasFractionalWeights |= weight != Math.rint(weight); + bufferIncomingCentroid(centroid.mean(), weight); + _originalFractionalBytes = null; + } + if (otherWeight != 0.0) { + _min = Math.min(_min, other.getMin()); + _max = Math.max(_max, other.getMax()); + } + } + + @Override + public void add(List<? extends TDigest> others) { + for (TDigest other : others) { + add(other); + } + } + + private void addAccumulator(PercentileTDigestAccumulator other) { + if (other == this) { + compress(); + if (_numCentroids == 0) { + return; + } + // Snapshot only self-merges; fresh fractional mass need not be representable by a legacy wire payload. + addCentroids(Arrays.copyOf(_centroidMeans, _numCentroids), Arrays.copyOf(_centroidWeights, _numCentroids), + _numCentroids, _min, _max, false); + return; + } + if (other._originalFractionalBytes != null && _compression == other._compression && 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(); + materializePendingSerializedTDigest(); + checkTotalWeight(_totalWeight + _incomingWeight + _numRawValues + other._totalWeight); + _hasFractionalWeights |= other._hasFractionalWeights; + preserveSerializedCapacity(other._serializedMainCapacity, other._serializedBufferCapacity); + bufferIncomingCentroids(other._centroidMeans, other._centroidWeights, other._numCentroids); + if (other._totalWeight > 0.0) { + _originalFractionalBytes = null; + } + _min = Math.min(_min, other._min); + _max = Math.max(_max, other._max); + } + + public void addSerializedTDigest(byte[] bytes) { + 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) { + requireMutable(); + if (prepareSerializedTDigest(input)) { + mergeSerializedTDigest(input, false); + } + } + + public void addSerializedTDigestDirect(SerializedTDigestInput input) { + requireMutable(); + if (prepareSerializedTDigest(input)) { + mergeSerializedTDigest(input, true); + } + } + + private boolean prepareSerializedTDigest(SerializedTDigestInput input) { + if (_pendingSerializedTDigest == null && hasNoInputs()) { + input.inspectMetadata(); + _hasFractionalWeights |= input._metadata.fractionalWeights(); + _pendingSerializedTDigest = input.retainBytes(); + _pendingSerializedMetadata = input._metadata; + if (input.hasUnrepresentableBoundaryMass() && _compression == input._metadata.compression()) { + // 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._metadata; + _legacyDegraded = true; + return false; + } + throw corruptedMutation(); + } + checkTotalWeight(_totalWeight + _incomingWeight + _numRawValues + input._metadata.totalWeight()); + 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; + } + preserveSerializedCapacity(input._metadata.mainCapacity(), input._metadata.bufferCapacity()); + 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); + } + _min = Math.min(_min, input._metadata.min()); + _max = Math.max(_max, input._metadata.max()); + } + + @Override + public void compress() { + if (_legacyDegraded) { + return; + } + if (_pendingSerializedTDigest != null || _numRawValues > 0 || _numIncomingCentroids > 0 + || !_publiclyCompressed && _numCentroids > 0) { + _serializedBytesForWrite = null; + } + materializePendingSerializedTDigest(); + if (_numRawValues > 0 || _numIncomingCentroids > 0) { + // MergingDigest 3.3 combines pending values and existing centroids directly at public compression. Flushing at + // working compression first would add an extra lossy merge pass and materially increase reducer rank error. + flush(_compression); + _publiclyCompressed = true; + } else if (!_publiclyCompressed && _numCentroids > 0) { + recompressCentroids(_compression); + _publiclyCompressed = true; + } + } + + @Override + public long size() { + return (long) getTotalWeight(); + } + + @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])) { + return (weightSoFar + _centroidWeights[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 (quantile == 0.0) { + return _min; + } + if (quantile == 1.0) { + return _max; + } + if (_numCentroids == 1) { + return _centroidMeans[0]; + } + + 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 = TDigestUtils.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() == TDigestUtils.VERBOSE_CENTROID_SIZE + ? encoded.getDouble() : encoded.getFloat(); + double mean = metadata.centroidSize() == TDigestUtils.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; + } + + /// Returns the length of the [#serialize()] bytes rather than the raw verbose byte size, so generic `TDigest` + /// serializers ([#byteSize()] followed by [#asBytes(ByteBuffer)]) emit the mixed-version-compatible encoding. + /// Prepared bytes are cached until the next write or mutation, avoiding a second encoding pass for that pattern. + /// Emitting raw verbose bytes here could produce a digest with more centroids than an old reader of the same + /// compression allocates, which it rejects with [ArrayIndexOutOfBoundsException]. + @Override + public int byteSize() { + if (_legacyDegraded || _originalFractionalBytes != null) { + return _legacyDegraded ? _pendingSerializedTDigest.length : _originalFractionalBytes.length; + } + if (_serializedBytesForWrite == null) { + _serializedBytesForWrite = serialize(); + } + return _serializedBytesForWrite.length; + } + + /// Bounds both compatible encodings and a widened verbose view without flushing. Buffered and fractional inputs + /// can reserve more space than their eventual bytes; retained capacities need not fit default compression size. + @Override + public int maxSerializedByteSize() { + if (_legacyDegraded) { + return _pendingSerializedTDigest.length; + } + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.length; + } + long centroidCount = getCentroidCountUpperBound(); + if (!_hasFractionalWeights) { + centroidCount = (long) Math.min(centroidCount, Math.ceil(getTotalWeight())); + } + int maximumBytes = getMaxVerboseByteSize(centroidCount); + return _pendingSerializedTDigest == null ? maximumBytes : Math.max(maximumBytes, _pendingSerializedTDigest.length); + } + + private static int getMaxVerboseByteSize(long centroidCount) { + return Math.toIntExact(Math.addExact(TDigestUtils.VERBOSE_HEADER_SIZE, + Math.multiplyExact(TDigestUtils.VERBOSE_CENTROID_SIZE, centroidCount))); + } + + /// Returns the small-encoding size unless float narrowing would overflow, in which case verbose bytes are used. + /// Historical degraded state retains its original encoding and rejects mutation. + @Override + public int smallByteSize() { + if (_legacyDegraded || _originalFractionalBytes != null) { + return _legacyDegraded ? _pendingSerializedTDigest.length : _originalFractionalBytes.length; + } + compress(); + normalizeBoundaryCentroids(); + checkFractionalBoundaryEncoding(); + checkCapacityPreservingCentroidCount(); + if (!canUseSmallEncoding()) { + return byteSize(); + } + return Math.addExact(TDigestUtils.SMALL_HEADER_SIZE, + Math.multiplyExact(TDigestUtils.SMALL_CENTROID_SIZE, _numCentroids)); + } + + /// Writes the [#serialize()] bytes; see [#byteSize()]. Bytes are always written in big-endian order (the t-digest + /// wire order) regardless of the destination buffer's byte order, unlike the library's `asBytes`. + @Override + public void asBytes(ByteBuffer buffer) { + byte[] bytes = _serializedBytesForWrite; + _serializedBytesForWrite = null; + buffer.put(bytes != null ? bytes : serialize()); + } + + /// Writes the representation described by [#smallByteSize()]; readers must inspect the leading encoding identifier. + @Override + public void asSmallBytes(ByteBuffer buffer) { + if (_legacyDegraded || _originalFractionalBytes != null) { + buffer.put(serialize()); + return; + } + compress(); + normalizeBoundaryCentroids(); + checkFractionalBoundaryEncoding(); + checkCapacityPreservingCentroidCount(); + if (canUseSmallEncoding()) { + buffer.put(toCapacityPreservingBytes()); + } else { + asBytes(buffer); + } + } + + @Override + public int centroidCount() { + if (_pendingSerializedTDigest != null) { + return !_legacyDegraded && _pendingSerializedMetadata.totalWeight() == 0.0 + ? 0 : _pendingSerializedMetadata.centroidCount(); + } + 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; + } + + public byte[] serialize() { + if (_legacyDegraded) { + return _pendingSerializedTDigest.clone(); + } + if (_originalFractionalBytes != null) { + return _originalFractionalBytes.clone(); + } + if (_pendingSerializedTDigest != null) { + if (_pendingSerializedMetadata.weightedBoundaries() || _pendingSerializedMetadata.hasZeroWeightCentroids() + || !Double.isNaN(_pendingSerializedMetadata.recoveredInfinityMean())) { + materializePendingSerializedTDigest(); + } else { + if (ByteBuffer.wrap(_pendingSerializedTDigest).getInt() == TDigestUtils.VERBOSE_ENCODING) { + byte[] serialized = TDigestUtils.makeLegacyCompatible(_pendingSerializedTDigest, _pendingSerializedMetadata); + return serialized == _pendingSerializedTDigest ? serialized.clone() : serialized; + } + return _pendingSerializedTDigest.clone(); + } + } + compress(); + normalizeBoundaryCentroids(); + checkFractionalBoundaryEncoding(); Review Comment: **[severe] Merging a legacy sub-1.0 fractional-boundary digest with healthy mass makes the group intermediate unserializable, failing the whole query.** `mergeSorted` pins the first/last inputs as singletons (`inputIndex == 1 || inputIndex == numInputs - 1 -> canMerge = false`), so a historical boundary centroid with weight in (0,1) that remains the global min/max keeps its fractional weight through every merge. `normalizeBoundaries` only repairs weights > 1.0, and once healthy mass is merged `_originalFractionalBytes` is cleared (`prepareSerializedTDigest`), so `serialize()` falls through to `checkFractionalBoundaryEncoding()` -> `checkLegacyBoundaryWeights` and throws `"Cannot serialize fractional TDigest boundary mass for legacy readers"`. Leaf servers always serialize intermediate results (`AggregationResultsBlock`/`GroupByResultsBlock.serializeIntermediateResult`), and nothing in the combine/reduce stack catches `IllegalArgumentException`, so one legacy fractional-boundary row plus healthy rows in the same group turns a previously-answerable query into a hard error (base had no such check). The only surviving topology is when the fractional payload and healthy mass live on different servers and meet only at broker `merge()` (no re-serialization afterwards). `PercentileTDigestAccumulatorTest` covers fresh-fractional throws and read-only retention, but not this merged shape; consider degrading (retain-bytes/NaN) or repairing instead of throwing when the fractional boundary comes from retained historical state, and add a test merging a 0.5-boundary payload with healthy mass followed by `serialize()`. -- 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]
