gortiz commented on code in PR #8979: URL: https://github.com/apache/pinot/pull/8979#discussion_r982188923
########## pinot-core/src/main/java/org/apache/pinot/core/operator/query/LinearSelectionOrderByOperator.java: ########## @@ -0,0 +1,476 @@ +/** + * 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.core.operator.query; + +import com.google.common.base.Preconditions; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Collections; +import java.util.Comparator; +import java.util.Iterator; +import java.util.List; +import java.util.PriorityQueue; +import java.util.function.IntFunction; +import java.util.function.Supplier; +import javax.annotation.Nullable; +import org.apache.pinot.common.request.context.ExpressionContext; +import org.apache.pinot.common.request.context.OrderByExpressionContext; +import org.apache.pinot.common.utils.DataSchema; +import org.apache.pinot.core.common.BlockValSet; +import org.apache.pinot.core.common.Operator; +import org.apache.pinot.core.common.RowBasedBlockValueFetcher; +import org.apache.pinot.core.operator.BaseOperator; +import org.apache.pinot.core.operator.ExecutionStatistics; +import org.apache.pinot.core.operator.blocks.TransformBlock; +import org.apache.pinot.core.operator.blocks.results.SelectionResultsBlock; +import org.apache.pinot.core.operator.transform.TransformOperator; +import org.apache.pinot.core.operator.transform.TransformResultMetadata; +import org.apache.pinot.core.query.request.context.QueryContext; +import org.apache.pinot.core.query.selection.SelectionOperatorUtils; +import org.apache.pinot.segment.spi.IndexSegment; +import org.roaringbitmap.RoaringBitmap; + + +/** + * A selection Operator used when the first expressions in the order by are identifier expressions of columns that are + * already sorted (either ascendingly or descendingly), even if the tail of order by expressions are not sorted. + * + * ie: SELECT ... FROM Table WHERE predicates ORDER BY sorted_column DESC LIMIT 10 OFFSET 5 + * or: SELECT ... FROM Table WHERE predicates ORDER BY sorted_column, not_sorted LIMIT 10 OFFSET 5 + * but not SELECT ... FROM Table WHERE predicates ORDER BY not_sorted, sorted_column LIMIT 10 OFFSET 5 + * + * Operators that derives from this class are going to have an almost linear cost instead of the usual NlogN when actual + * sorting must be done, where N is the number of rows in the segment. + * There is a degraded scenario when the cost is actually NlogL (where L is the limit of the query) when all the first L + * rows have the exact same value for the prefix of the sorted columns. Even in that case, L should be quite smaller + * than N, so this implementation is algorithmically better than the normal solution. + */ +public abstract class LinearSelectionOrderByOperator extends BaseOperator<SelectionResultsBlock> { + protected final IndexSegment _indexSegment; + + protected final boolean _nullHandlingEnabled; + // Deduped order-by expressions followed by output expressions from SelectionOperatorUtils.extractExpressions() + protected final List<ExpressionContext> _expressions; + protected final List<ExpressionContext> _alreadySorted; + protected final List<ExpressionContext> _toSort; + + protected final TransformOperator _transformOperator; + protected final List<OrderByExpressionContext> _orderByExpressions; + protected final TransformResultMetadata[] _expressionsMetadata; + protected final int _numRowsToKeep; + private final Supplier<ListBuilder> _listBuilderSupplier; + protected boolean _used = false; + protected Comparator<Object[]> _comparator; + + /** + * + * @param expressions order by expressions must be at the head of the list. + * @param sortedExpr How many expressions at the head of the expression list are going to be considered sorted by + * {@link #fetch(Supplier<ListBuilder>)} + */ + public LinearSelectionOrderByOperator(IndexSegment indexSegment, QueryContext queryContext, + List<ExpressionContext> expressions, TransformOperator transformOperator, + int sortedExpr) { + _indexSegment = indexSegment; + _nullHandlingEnabled = queryContext.isNullHandlingEnabled(); + _expressions = expressions; + _transformOperator = transformOperator; + + _orderByExpressions = queryContext.getOrderByExpressions(); + assert _orderByExpressions != null; + int numOrderByExpressions = _orderByExpressions.size(); + + _alreadySorted = expressions.subList(0, sortedExpr); + _toSort = expressions.subList(sortedExpr, numOrderByExpressions); + + _expressionsMetadata = new TransformResultMetadata[_expressions.size()]; + for (int i = 0; i < _expressionsMetadata.length; i++) { + ExpressionContext expression = _expressions.get(i); + _expressionsMetadata[i] = _transformOperator.getResultMetadata(expression); + } + + _numRowsToKeep = queryContext.getOffset() + queryContext.getLimit(); + + if (_toSort.isEmpty()) { + int expectedSize = Math.min(SelectionOperatorUtils.MAX_ROW_HOLDER_INITIAL_CAPACITY, _numRowsToKeep); + _listBuilderSupplier = () -> new TotallySortedListBuilder(expectedSize); + } else { + int maxSize = Math.min(SelectionOperatorUtils.MAX_ROW_HOLDER_INITIAL_CAPACITY, _numRowsToKeep * 2); + + Comparator<Object[]> sortedComparator = OrderByComparatorFactory.getComparator(_orderByExpressions, + _expressionsMetadata, true, _nullHandlingEnabled, 0, sortedExpr); + Comparator<Object[]> unsortedComparator = OrderByComparatorFactory.getComparator(_orderByExpressions, + _expressionsMetadata, true, _nullHandlingEnabled, sortedExpr, numOrderByExpressions); + _listBuilderSupplier = () -> new PartiallySortedListBuilder(maxSize, sortedComparator, unsortedComparator); + } + + _comparator = + OrderByComparatorFactory.getComparator(_orderByExpressions, _expressionsMetadata, false, _nullHandlingEnabled); + } + + @Override + public IndexSegment getIndexSegment() { + return _indexSegment; + } + + @Override + public ExecutionStatistics getExecutionStatistics() { + long numEntriesScannedInFilter = _transformOperator.getExecutionStatistics().getNumEntriesScannedInFilter(); + int numTotalDocs = _indexSegment.getSegmentMetadata().getTotalDocs(); + return new ExecutionStatistics(getNumDocsScanned(), numEntriesScannedInFilter, getNumEntriesScannedPostFilter(), + numTotalDocs); + } + + protected IntFunction<Object[]> fetchBlock(TransformBlock transformBlock, BlockValSet[] blockValSets) { + int numExpressions = _expressions.size(); + + for (int i = 0; i < numExpressions; i++) { + ExpressionContext expression = _expressions.get(i); + blockValSets[i] = transformBlock.getBlockValueSet(expression); + } + RowBasedBlockValueFetcher blockValueFetcher = new RowBasedBlockValueFetcher(blockValSets); + + if (!_nullHandlingEnabled) { + return blockValueFetcher::getRow; + } + RoaringBitmap[] nullBitmaps = new RoaringBitmap[numExpressions]; + for (int i = 0; i < numExpressions; i++) { + nullBitmaps[i] = blockValSets[i].getNullBitmap(); + } + return (docId) -> { + Object[] row = blockValueFetcher.getRow(docId); + for (int colId = 0; colId < nullBitmaps.length; colId++) { + if (nullBitmaps[colId] != null && nullBitmaps[colId].contains(docId)) { + row[colId] = null; + } + } + return row; + }; + } + + protected abstract long getNumEntriesScannedPostFilter(); + + protected abstract int getNumDocsScanned(); + + /** + * Returns a list of rows sorted that: + * <ul> + * <li>At least contains all the rows that fulfill the predicate</li> + * <li>Rows are sorted in a way that is compatible with the given list builder supplier</li> + * </ul> + * + * That means that the result may contain more rows than required. + * + * @param listBuilderSupplier a {@link ListBuilder} supplier that should be used to create the result. Each time is + * called a new {@link ListBuilder} will be returned. All returned instances use the same + * comparator logic. + */ + protected abstract List<Object[]> fetch(Supplier<ListBuilder> listBuilderSupplier); + + @Override + public List<Operator> getChildOperators() { + return Collections.singletonList(_transformOperator); + } + + protected abstract String getExplainName(); + + @Override + public String toExplainString() { + StringBuilder sb = new StringBuilder(getExplainName()); + + sb.append("(sortedList: "); + concatList(sb, _alreadySorted); + + sb.append(", unsortedList: "); + concatList(sb, _toSort); + + sb.append(", rest: "); + concatList(sb, _expressions.subList(_alreadySorted.size() + _toSort.size(), _expressions.size())); + + sb.append(')'); + return sb.toString(); + } + + private void concatList(StringBuilder sb, List<?> list) { + sb.append('('); + Iterator<?> it = list.iterator(); + if (it.hasNext()) { + sb.append(it.next()); + while (it.hasNext()) { + sb.append(", ").append(it.next()); + } + } + sb.append(')'); + } + + @Override + protected SelectionResultsBlock getNextBlock() { + Preconditions.checkState(!_used, "nextBlock was called more than once"); + _used = true; + List<Object[]> list = fetch(_listBuilderSupplier); + + DataSchema dataSchema = createDataSchema(); + + if (list.size() > _numRowsToKeep) { + list = new ArrayList<>(list.subList(0, _numRowsToKeep)); + } + + SelectionResultsBlock resultsBlock = new SelectionResultsBlock(dataSchema, list, _comparator); + + return resultsBlock; + } + + protected DataSchema createDataSchema() { + int numExpressions = _expressions.size(); + + // Create the data schema + String[] columnNames = new String[numExpressions]; + DataSchema.ColumnDataType[] columnDataTypes = new DataSchema.ColumnDataType[numExpressions]; + for (int i = 0; i < columnNames.length; i++) { + columnNames[i] = _expressions.get(i).toString(); + } + for (int i = 0; i < numExpressions; i++) { + TransformResultMetadata expressionMetadata = _expressionsMetadata[i]; + columnDataTypes[i] = + DataSchema.ColumnDataType.fromDataType(expressionMetadata.getDataType(), expressionMetadata.isSingleValue()); + } + return new DataSchema(columnNames, columnDataTypes); + } + + /** + * A private class used to build a sorted list by adding partially sorted data. + * + * Specifically, this class has been designed to receive successive calls to {@link #add(Object[])} follow by a single + * call to {@link #build()}. Once this method is called, the behavior of this object is undefined and therefore it + * should not be used. + * + * Rows must be inserted in ascending order accordingly to the partial order specified by a comparator. + * This comparator will define <i>partitions</i> of rows. All the rows in the same partition same are considered equal + * by that comparator. + * + * Some implementations, like {@link PartiallySortedListBuilder} accepts a second comparator that is used to sort rows + * inside each partition. + * When calling {@link #add(Object[])} with a row that doesn't belong to the current partition, a new one is started + * and the previous one is sorted. Therefore, this object maintains the invariant that at any moment there are at + * least {@link #sortedSize()} elements that are completely sorted, which means that no matter which elements are + * added after that, these elements are going to be smallest. + */ + protected interface ListBuilder { + + /** + * Adds the given row to this object. The new column must be equal or higher than the partition comparator. + * + * @param row The row to add. The values of the already sorted columns must be equal or higher than the last added + * row, if any. + * @return true if and only if the previous partition was closed. + */ + boolean add(Object[] row); + + /** + * Builds the sorted list. The returned list will be sorted, even if a previous call t o {@link #sortedSize()} + * returned a number lower than the number of elements that have been added. + * + * Once this method is called, the builder should not be used. There is no guaranteed on whether + * {@link #sortedSize()} is updated or not. + */ + List<Object[]> build(); + + /** + * How many elements are actually sorted. This number is lower or equal to the number of elements that has been + * added. + */ + int sortedSize(); + } + + /** + * This is the faster {@link ListBuilder} but also the most restrictive one. It can only be used when data is inserted + * in total order. Therefore it cannot be used to implement order-by queries where there is at least one expression + * that is not sorted. In such case {@link PartiallySortedListBuilder} should be used. + * + * This implementation is just a wrapper over an ArrayList and therefore the average costs of its methods is constant. + */ + protected static class TotallySortedListBuilder implements ListBuilder { + private final ArrayList<Object[]> _list; + + public TotallySortedListBuilder(int expectedSize) { + _list = new ArrayList<>(expectedSize); + } + + @Override + public boolean add(Object[] row) { + _list.add(row); + return true; + } + + @Override + public List<Object[]> build() { + return _list; + } + + @Override + public int sortedSize() { + return _list.size(); + } + } + + /** + * This is the more versatile {@link ListBuilder} that requires two comparators: The first defines the order on which + * {@link #add(Object[])} is called and defines the partitions. The second one is used to sort rows in the same + * partition. This class can be used to implement order-by queries that include one or more not sorted expressions. + * In cases where all expressions are sorted, {@link TotallySortedListBuilder} should be used because its performance + * is better. + * + * In this implementation the {@link #add(Object[])} method has a log(P) cost, where P is the number of elements in + * each partition. In normal cases we can consider P a constant. In some degenerated cases like having an ordering by + * a sorted column whose values are all equal, P may tend to the number of elements in the segment, which would make + * the cost of this method log(N). + * + * This class is optimized for scenarios where the partitions are smaller than the limit. In that case, the partition + * is only sorted once it is closed (aka when a new element for the other partition is found), so each insertion has + * an average cost of log(P) (where P-1 insertions are cost O(1) and the last one is O(PlogP)). In cases where at + * least one partition is larger than the limit, this implementation delegates the last partition to a + * {@link PriorityQueue}. + */ + private static class PartiallySortedListBuilder implements ListBuilder { + private final ArrayList<Object[]> _rows; + /** + * The comparator that defines the partitions and the one that impose in which order add has to be called. + */ + private final Comparator<Object[]> _partitionComparator; + /** + * The comparator that sorts different rows on each partition. + */ + private final Comparator<Object[]> _unsortedComparator; + @Nullable + private Object[] _lastPartitionRow; + /** + * How many elements are already sorted. This is also the index where the last (and still not sorted) partition + * starts or should start. + */ + private int _sortedPrefixSize; + private final int _maxSize; + private PriorityQueue<Object[]> _queue = null; + private int _maxPriorityQueueValues; + + public PartiallySortedListBuilder(int maxSize, Comparator<Object[]> partitionComparator, + Comparator<Object[]> unsortedComparator) { + Preconditions.checkArgument(maxSize > 0, "Max size must be greater than 0"); + _maxSize = maxSize; + _rows = new ArrayList<>(maxSize); + _partitionComparator = partitionComparator; + _unsortedComparator = unsortedComparator; + } + + @Override + public boolean add(Object[] row) { + boolean result; + if (_lastPartitionRow == null) { + _sortedPrefixSize = 0; + _lastPartitionRow = row; + _rows.add(row); + return false; + } else { + int cmp = _partitionComparator.compare(row, _lastPartitionRow); + if (cmp < 0) { + throw new IllegalArgumentException("Row " + Arrays.toString(row) + " is lower than previously added row" + + Arrays.toString(_lastPartitionRow)); + } + + if (_queue != null) { + if (cmp > 0) { + // the new row is in a greater partition, just don't add it and continue + return true; + } + // otherwise they are equal by the already sorted columns. Let's check the other columns and evict the higher + // also, we have the higher value at the last position + SelectionOperatorUtils.addToPriorityQueue(row, _queue, _maxPriorityQueueValues); + return false; + } else { + if (_rows.size() == _maxSize) { // we already have _maxSize results + if (cmp > 0) { + // and the new one is higher, so we don't need to add it + sortLastPartition(); + return true; + } + // and the new one may be smaller, so this is going to be the last partition and we need to prepare the + // priority queue + buildPriorityQueue(); + SelectionOperatorUtils.addToPriorityQueue(row, _queue, _maxPriorityQueueValues); + return false; + } + if (cmp > 0) { + sortLastPartition(); + _lastPartitionRow = row; + result = true; + } else { + result = false; + } + _rows.add(row); + return result; + } + } + } + + private void buildPriorityQueue() { + _maxPriorityQueueValues = _maxSize - _sortedPrefixSize; + _queue = new PriorityQueue<>(_maxPriorityQueueValues, _unsortedComparator.reversed()); + for (int i = 0; i < _maxPriorityQueueValues; i++) { + int lastIdx = _rows.size() - 1; + _queue.add(_rows.get(lastIdx)); + _rows.remove(lastIdx); + } + } + + void sortLastPartition() { Review Comment: So your idea is that `ListBuilder.build()` invariant should be that elements are ordered by partition, but not by the not already sorted columns. In the current implementation, I think that makes sense. I'm going to change it. But I think we should think at longer term. My original idea was more ambitious than that. The idea was to sort each block in the map phase (therefore in parallel) so the reduce/combine phase (that is called by a single thread) should have a linear cost. Also, by applying the sorting in the map phase we can use the local knowledge derived from sorted columns and make the comparison faster by ignoring the fields that are already sorted in the partition. -- 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: commits-unsubscr...@pinot.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@pinot.apache.org For additional commands, e-mail: commits-h...@pinot.apache.org
