jpountz commented on a change in pull request #101: URL: https://github.com/apache/lucene/pull/101#discussion_r627389726
########## File path: lucene/core/src/java/org/apache/lucene/search/BlockMaxMaxscoreScorer.java ########## @@ -0,0 +1,339 @@ +/* + * 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.lucene.search; + +import static org.apache.lucene.search.ScorerUtil.costWithMinShouldMatch; + +import java.io.IOException; +import java.util.*; + +/** Scorer implementing Block-Max Maxscore algorithm */ +public class BlockMaxMaxscoreScorer extends Scorer { + private final ScoreMode scoreMode; + private final int scalingFactor; + + // current doc ID of the leads + private int doc; + + // doc id boundary that all scorers maxScore are valid + private int upTo = -1; + + // heap of scorers ordered by doc ID + private final DisiPriorityQueue essentialsScorers; + + // list of scorers whose sum of maxScore is less than minCompetitiveScore, ordered by maxScore + private final List<DisiWrapper> nonEssentialScorers; + + // sum of max scores of scorers in nonEssentialScorers list + private long nonEssentialMaxScoreSum; + + // sum of score of scorers in essentialScorers list that are positioned on matching doc + private long matchedDocScoreSum; + + private long cost; + + private final MaxScoreSumPropagator maxScoreSumPropagator; + + private final List<Scorer> scorers; + + // scaled min competitive score + private long minCompetitiveScore = 0; + + /** + * Constructs a Scorer + * + * @param weight The weight to be used. + * @param scorers The sub scorers this Scorer should iterate on for optional clauses + * @param scoreMode The scoreMode + */ + public BlockMaxMaxscoreScorer(Weight weight, List<Scorer> scorers, ScoreMode scoreMode) + throws IOException { + super(weight); + assert scoreMode == ScoreMode.TOP_SCORES; + + this.scoreMode = scoreMode; + this.doc = -1; + this.scorers = scorers; + this.cost = + costWithMinShouldMatch( + scorers.stream().map(Scorer::iterator).mapToLong(DocIdSetIterator::cost), + scorers.size(), + 1); + + essentialsScorers = new DisiPriorityQueue(scorers.size()); + nonEssentialScorers = new LinkedList<>(); + + scalingFactor = WANDScorer.getScalingFactor(scorers); + maxScoreSumPropagator = new MaxScoreSumPropagator(scorers); + + for (Scorer scorer : scorers) { + nonEssentialScorers.add(new DisiWrapper(scorer)); + } + } + + @Override + public DocIdSetIterator iterator() { + return TwoPhaseIterator.asDocIdSetIterator(twoPhaseIterator()); + } + + @Override + public TwoPhaseIterator twoPhaseIterator() { + DocIdSetIterator approximation = + new DocIdSetIterator() { + private long lastMinCompetitiveScore; + + @Override + public int docID() { + return doc; + } + + @Override + public int nextDoc() throws IOException { + return advance(doc + 1); + } + + @Override + public int advance(int target) throws IOException { + doAdvance(target); + + while (doc != DocIdSetIterator.NO_MORE_DOCS + && nonEssentialMaxScoreSum + matchedDocScoreSum < minCompetitiveScore) { + doAdvance(doc + 1); + } + + return doc; + } + + private void doAdvance(int target) throws IOException { + matchedDocScoreSum = 0; + // Find next smallest doc id that is larger than or equal to target from the essential + // scorers + + // If the next candidate doc id is still within interval boundary, + if (lastMinCompetitiveScore == minCompetitiveScore && target <= upTo) { + while (essentialsScorers.top().doc < target) { + DisiWrapper w = essentialsScorers.pop(); + w.doc = w.iterator.advance(target); + essentialsScorers.add(w); + } + + if (essentialsScorers.top().doc <= upTo) { + doc = essentialsScorers.top().doc; + + if (doc == NO_MORE_DOCS) { + return; + } + } else { + doc = upTo + 1; + } + } else { + lastMinCompetitiveScore = minCompetitiveScore; + // Next candidate doc id is above interval boundary, or minCompetitive has increased. + // Find next interval boundary. + // Block boundary alignment strategy is adapted from "Optimizing Top-k Document + // Retrieval Strategies for Block-Max Indexes" by Dimopoulos, Nepomnyachiy and Suel. + // Find the block interval boundary that is the minimum of all participating scorer's + // block boundary. Then run BMM within each interval. + updateUpToAndMaxScore(target); + + repartitionLists(); + + // maxScore of all scorers sum to less than minCompetitiveScore, no more result is + // available up to upTo + if (essentialsScorers.size() == 0) { + // current bound no long valid + doc = upTo + 1; + } else { + doc = essentialsScorers.top().doc; + + if (doc == NO_MORE_DOCS) { + return; + } + } + } + + for (DisiWrapper w : essentialsScorers) { + if (w.doc == doc && doc != NO_MORE_DOCS) { + matchedDocScoreSum += WANDScorer.scaleMaxScore(w.scorer.score(), scalingFactor); + } + } + } + + private void updateUpToAndMaxScore(int target) throws IOException { + while (essentialsScorers.size() > 0) { + nonEssentialScorers.add(essentialsScorers.pop()); + } + + // reset upTo + upTo = DocIdSetIterator.NO_MORE_DOCS; + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc < target) { + upTo = Math.min(w.scorer.advanceShallow(target), upTo); + } else { + upTo = Math.min(w.scorer.advanceShallow(w.doc), upTo); + } + } + assert target <= upTo; + + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc <= upTo) { + w.maxScore = WANDScorer.scaleMaxScore(w.scorer.getMaxScore(upTo), scalingFactor); + } else { + // This scorer won't be able to contribute to match for target, setting its maxScore + // to 0 so it goes into nonEssentialList + w.maxScore = 0; + } + if (w.doc < target) { + w.doc = w.iterator.advance(target); + } + } + } + + private void repartitionLists() { + Collections.sort(nonEssentialScorers, (w1, w2) -> (int) (w1.maxScore - w2.maxScore)); + + // Re-partition the scorers into non-essential list and essential list, as defined in + // the "Optimizing Top-k Document Retrieval Strategies for Block-Max Indexes" paper. + nonEssentialMaxScoreSum = 0; + for (int i = 0; i < nonEssentialScorers.size(); i++) { + DisiWrapper w = nonEssentialScorers.get(i); + if (nonEssentialMaxScoreSum + w.maxScore < minCompetitiveScore) { + nonEssentialMaxScoreSum += w.maxScore; + } else { + // the logic is a bit ugly here...but as soon as we find maxScore of scorers in + // non-essential list sum to above minCompetitiveScore, we move the rest of + // scorers + // into essential list + for (int j = nonEssentialScorers.size() - 1; j >= i; j--) { + essentialsScorers.add(nonEssentialScorers.remove(j)); + } + break; + } + } + } + + @Override + public long cost() { + // fixed at initialization + return cost; + } + }; + + return new TwoPhaseIterator(approximation) { + @Override + public boolean matches() throws IOException { + // The doc is a match when all scores sum above minCompetitiveScore + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc < doc) { + w.doc = w.iterator.advance(doc); + } + } + + if (matchedDocScoreSum >= minCompetitiveScore) { + return true; + } + + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc == doc) { + matchedDocScoreSum += WANDScorer.scaleMaxScore(w.scorer.score(), scalingFactor); + + if (matchedDocScoreSum >= minCompetitiveScore) { + return true; + } + } + } + + return false; + } Review comment: I don't think that implementing a TwoPhaseIterator this way helps, since we're computing scores once here to check whether the hit is competitive, and then again in `score()` if the hit is competitive. Let's return the approximation directly as `iterator()` and hits that don't have a competitive scores will naturally be ignored after `score()` is computed since their scores will be lower than the score of the top of the priority queue? ########## File path: lucene/core/src/java/org/apache/lucene/search/BlockMaxMaxscoreScorer.java ########## @@ -0,0 +1,339 @@ +/* + * 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.lucene.search; + +import static org.apache.lucene.search.ScorerUtil.costWithMinShouldMatch; + +import java.io.IOException; +import java.util.*; + +/** Scorer implementing Block-Max Maxscore algorithm */ +public class BlockMaxMaxscoreScorer extends Scorer { + private final ScoreMode scoreMode; + private final int scalingFactor; + + // current doc ID of the leads + private int doc; + + // doc id boundary that all scorers maxScore are valid + private int upTo = -1; + + // heap of scorers ordered by doc ID + private final DisiPriorityQueue essentialsScorers; + + // list of scorers whose sum of maxScore is less than minCompetitiveScore, ordered by maxScore + private final List<DisiWrapper> nonEssentialScorers; + + // sum of max scores of scorers in nonEssentialScorers list + private long nonEssentialMaxScoreSum; + + // sum of score of scorers in essentialScorers list that are positioned on matching doc + private long matchedDocScoreSum; + + private long cost; + + private final MaxScoreSumPropagator maxScoreSumPropagator; + + private final List<Scorer> scorers; + + // scaled min competitive score + private long minCompetitiveScore = 0; + + /** + * Constructs a Scorer + * + * @param weight The weight to be used. + * @param scorers The sub scorers this Scorer should iterate on for optional clauses + * @param scoreMode The scoreMode + */ + public BlockMaxMaxscoreScorer(Weight weight, List<Scorer> scorers, ScoreMode scoreMode) + throws IOException { + super(weight); + assert scoreMode == ScoreMode.TOP_SCORES; + + this.scoreMode = scoreMode; + this.doc = -1; + this.scorers = scorers; + this.cost = + costWithMinShouldMatch( + scorers.stream().map(Scorer::iterator).mapToLong(DocIdSetIterator::cost), + scorers.size(), + 1); + + essentialsScorers = new DisiPriorityQueue(scorers.size()); + nonEssentialScorers = new LinkedList<>(); + + scalingFactor = WANDScorer.getScalingFactor(scorers); + maxScoreSumPropagator = new MaxScoreSumPropagator(scorers); + + for (Scorer scorer : scorers) { + nonEssentialScorers.add(new DisiWrapper(scorer)); + } + } + + @Override + public DocIdSetIterator iterator() { + return TwoPhaseIterator.asDocIdSetIterator(twoPhaseIterator()); + } + + @Override + public TwoPhaseIterator twoPhaseIterator() { + DocIdSetIterator approximation = + new DocIdSetIterator() { + private long lastMinCompetitiveScore; + + @Override + public int docID() { + return doc; + } + + @Override + public int nextDoc() throws IOException { + return advance(doc + 1); + } + + @Override + public int advance(int target) throws IOException { + doAdvance(target); + + while (doc != DocIdSetIterator.NO_MORE_DOCS + && nonEssentialMaxScoreSum + matchedDocScoreSum < minCompetitiveScore) { + doAdvance(doc + 1); + } + + return doc; + } + + private void doAdvance(int target) throws IOException { + matchedDocScoreSum = 0; + // Find next smallest doc id that is larger than or equal to target from the essential + // scorers + + // If the next candidate doc id is still within interval boundary, + if (lastMinCompetitiveScore == minCompetitiveScore && target <= upTo) { + while (essentialsScorers.top().doc < target) { + DisiWrapper w = essentialsScorers.pop(); + w.doc = w.iterator.advance(target); + essentialsScorers.add(w); + } + + if (essentialsScorers.top().doc <= upTo) { + doc = essentialsScorers.top().doc; + + if (doc == NO_MORE_DOCS) { + return; + } + } else { + doc = upTo + 1; + } + } else { + lastMinCompetitiveScore = minCompetitiveScore; + // Next candidate doc id is above interval boundary, or minCompetitive has increased. + // Find next interval boundary. + // Block boundary alignment strategy is adapted from "Optimizing Top-k Document + // Retrieval Strategies for Block-Max Indexes" by Dimopoulos, Nepomnyachiy and Suel. + // Find the block interval boundary that is the minimum of all participating scorer's + // block boundary. Then run BMM within each interval. + updateUpToAndMaxScore(target); + + repartitionLists(); + + // maxScore of all scorers sum to less than minCompetitiveScore, no more result is + // available up to upTo + if (essentialsScorers.size() == 0) { + // current bound no long valid + doc = upTo + 1; + } else { + doc = essentialsScorers.top().doc; + + if (doc == NO_MORE_DOCS) { + return; + } + } + } + + for (DisiWrapper w : essentialsScorers) { + if (w.doc == doc && doc != NO_MORE_DOCS) { + matchedDocScoreSum += WANDScorer.scaleMaxScore(w.scorer.score(), scalingFactor); + } + } + } + + private void updateUpToAndMaxScore(int target) throws IOException { + while (essentialsScorers.size() > 0) { + nonEssentialScorers.add(essentialsScorers.pop()); + } + + // reset upTo + upTo = DocIdSetIterator.NO_MORE_DOCS; + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc < target) { + upTo = Math.min(w.scorer.advanceShallow(target), upTo); + } else { + upTo = Math.min(w.scorer.advanceShallow(w.doc), upTo); + } + } + assert target <= upTo; + + for (DisiWrapper w : nonEssentialScorers) { + if (w.doc <= upTo) { + w.maxScore = WANDScorer.scaleMaxScore(w.scorer.getMaxScore(upTo), scalingFactor); + } else { + // This scorer won't be able to contribute to match for target, setting its maxScore + // to 0 so it goes into nonEssentialList + w.maxScore = 0; + } + if (w.doc < target) { + w.doc = w.iterator.advance(target); + } Review comment: I wouldn't expect that we would have to advance scorers when updating `upTo`? -- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
