use of com.yahoo.search.predicate.index.conjunction.IndexableFeatureConjunction in project vespa by vespa-engine.
the class PredicateTreeAnnotator method assignIntervalLabels.
/**
* Visits the predicate tree in depth-first order and assigns intervals for features in
* {@link com.yahoo.document.predicate.FeatureSet} and {@link com.yahoo.document.predicate.FeatureRange}.
*/
private static void assignIntervalLabels(Predicate predicate, int begin, int end, boolean isNegated, AnnotatorContext context) {
// Otherwise, conjunctions and disjunctions must be switched if negated (De Morgan's law).
if (predicate instanceof Conjunction) {
List<Predicate> children = ((Conjunction) predicate).getOperands();
int current = begin;
for (int i = 0; i < children.size(); i++) {
Predicate child = children.get(i);
int subTreeSize = context.subTreeSizes.get(child);
if (i == children.size() - 1) {
// Last child (and sometimes the only one)
assignIntervalLabels(child, current, end, isNegated, context);
// No need to update/touch current since this is the last child.
} else if (i == 0) {
// First child
int next = context.leftNodeLeaves + subTreeSize + 1;
assignIntervalLabels(child, current, next - 1, isNegated, context);
current = next;
} else {
// Middle children
int next = current + subTreeSize;
assignIntervalLabels(child, current, next - 1, isNegated, context);
current = next;
}
}
} else if (predicate instanceof FeatureConjunction) {
// Register FeatureConjunction as it was a FeatureSet with a single child.
// Note: FeatureConjunction should never be negated as AndOrSimplifier will push negations down to
// the leafs (FeatureSets).
int zStarEnd = isNegated ? calculateZStarIntervalEnd(end, context) : end;
IndexableFeatureConjunction indexable = new IndexableFeatureConjunction((FeatureConjunction) predicate);
int interval = Interval.fromBoundaries(begin, zStarEnd);
context.featureConjunctions.computeIfAbsent(indexable, (k) -> new ArrayList<>()).add(interval);
if (isNegated) {
registerZStarInterval(begin, end, zStarEnd, context);
}
context.leftNodeLeaves += 1;
} else if (predicate instanceof Disjunction) {
// the values will be same as that of the parent OR node
for (Predicate child : ((Disjunction) predicate).getOperands()) {
assignIntervalLabels(child, begin, end, isNegated, context);
}
} else if (predicate instanceof FeatureSet) {
FeatureSet featureSet = (FeatureSet) predicate;
int zStarEnd = isNegated ? calculateZStarIntervalEnd(end, context) : end;
for (String value : featureSet.getValues()) {
long featureHash = Feature.createHash(featureSet.getKey(), value);
int interval = Interval.fromBoundaries(begin, zStarEnd);
registerFeatureInterval(featureHash, interval, context.intervals);
}
if (isNegated) {
registerZStarInterval(begin, end, zStarEnd, context);
}
context.leftNodeLeaves += 1;
} else if (predicate instanceof Negation) {
assignIntervalLabels(((Negation) predicate).getOperand(), begin, end, !isNegated, context);
} else if (predicate instanceof FeatureRange) {
FeatureRange featureRange = (FeatureRange) predicate;
int zStarEnd = isNegated ? calculateZStarIntervalEnd(end, context) : end;
int interval = Interval.fromBoundaries(begin, zStarEnd);
for (RangePartition partition : featureRange.getPartitions()) {
long featureHash = PredicateHash.hash64(partition.getLabel());
registerFeatureInterval(featureHash, interval, context.intervals);
}
for (RangeEdgePartition edgePartition : featureRange.getEdgePartitions()) {
long featureHash = PredicateHash.hash64(edgePartition.getLabel());
IntervalWithBounds intervalWithBounds = new IntervalWithBounds(interval, (int) edgePartition.encodeBounds());
registerFeatureInterval(featureHash, intervalWithBounds, context.intervalsWithBounds);
}
if (isNegated) {
registerZStarInterval(begin, end, zStarEnd, context);
}
context.leftNodeLeaves += 1;
} else {
throw new UnsupportedOperationException("Cannot handle predicate of type " + predicate.getClass().getSimpleName());
}
}
use of com.yahoo.search.predicate.index.conjunction.IndexableFeatureConjunction in project vespa by vespa-engine.
the class PredicateTreeAnnotatorTest method require_that_featureconjunctions_are_registered_and_given_an_interval.
@Test
public void require_that_featureconjunctions_are_registered_and_given_an_interval() {
Predicate p = and(or(range("key", partition("key=10-19"), partition("key=20-29"), edgePartition("key=0", 5, 10, 20), edgePartition("key=30", 0, 0, 3)), conj(not(feature("keyA").inSet("C")), feature("keyB").inSet("D"))), feature("foo").inSet("bar"));
PredicateTreeAnnotations r = PredicateTreeAnnotator.createPredicateTreeAnnotations(p);
assertEquals(2, r.minFeature);
assertEquals(3, r.intervalEnd);
assertEquals(3, r.intervalMap.size());
assertEquals(2, r.boundsMap.size());
assertEquals(1, r.featureConjunctions.size());
Map.Entry<IndexableFeatureConjunction, List<Integer>> entry = r.featureConjunctions.entrySet().iterator().next();
assertEquals(1, entry.getValue().size());
assertEquals(0b1_0000000000000010, entry.getValue().get(0).longValue());
}
use of com.yahoo.search.predicate.index.conjunction.IndexableFeatureConjunction in project vespa by vespa-engine.
the class PredicateIndexBuilder method indexDocumentConjunctions.
private void indexDocumentConjunctions(int docId, Map<IndexableFeatureConjunction, List<Integer>> featureConjunctions) {
for (Map.Entry<IndexableFeatureConjunction, List<Integer>> e : featureConjunctions.entrySet()) {
IndexableFeatureConjunction fc = e.getKey();
List<Integer> intervals = e.getValue();
Posting posting = new Posting(docId, intervalStoreBuilder.insert(intervals));
conjunctionIntervalIndexBuilder.insert(fc.id, posting);
conjunctionIndexBuilder.indexConjunction(fc);
}
}
use of com.yahoo.search.predicate.index.conjunction.IndexableFeatureConjunction in project vespa by vespa-engine.
the class PredicateTreeAnalyzer method aggregatePredicateStatistics.
// First analysis pass. Traverses tree in depth-first order. Determines the sub-tree sizes and counts the occurrences
// of each feature (used by min-feature calculation in second pass).
// Returns the size of the analyzed subtree.
private static int aggregatePredicateStatistics(Predicate predicate, boolean isNegated, AnalyzerContext context) {
if (predicate instanceof Negation) {
return aggregatePredicateStatistics(((Negation) predicate).getOperand(), !isNegated, context);
} else if (predicate instanceof Conjunction) {
return ((Conjunction) predicate).getOperands().stream().mapToInt(child -> {
int size = aggregatePredicateStatistics(child, isNegated, context);
context.subTreeSizes.put(child, size);
return size;
}).sum();
} else if (predicate instanceof FeatureConjunction) {
if (isNegated) {
context.hasNegationPredicate = true;
return 2;
}
// Count the number of identical feature conjunctions - use the id from IndexableFeatureConjunction as key
IndexableFeatureConjunction ifc = new IndexableFeatureConjunction((FeatureConjunction) predicate);
incrementOccurrence(context.conjunctionOccurrences, ifc.id);
// Handled as leaf in interval algorithm - count a single child
return 1;
} else if (predicate instanceof Disjunction) {
return ((Disjunction) predicate).getOperands().stream().mapToInt(child -> aggregatePredicateStatistics(child, isNegated, context)).sum();
} else if (predicate instanceof FeatureSet) {
if (isNegated) {
context.hasNegationPredicate = true;
return 2;
} else {
FeatureSet featureSet = (FeatureSet) predicate;
for (String value : featureSet.getValues()) {
incrementOccurrence(context.featureOccurrences, Feature.createHash(featureSet.getKey(), value));
}
return 1;
}
} else if (predicate instanceof FeatureRange) {
if (isNegated) {
context.hasNegationPredicate = true;
return 2;
} else {
incrementOccurrence(context.featureOccurrences, PredicateHash.hash64(((FeatureRange) predicate).getKey()));
return 1;
}
} else {
throw new UnsupportedOperationException("Cannot handle predicate of type " + predicate.getClass().getSimpleName());
}
}
use of com.yahoo.search.predicate.index.conjunction.IndexableFeatureConjunction in project vespa by vespa-engine.
the class PredicateTreeAnalyzer method findMinFeature.
// Second analysis pass. Traverses tree in depth-first order. Determines the min-feature value.
private static double findMinFeature(Predicate predicate, boolean isNegated, AnalyzerContext context) {
if (predicate instanceof Conjunction) {
// Sum of children values.
return ((Conjunction) predicate).getOperands().stream().mapToDouble(child -> findMinFeature(child, isNegated, context)).sum();
} else if (predicate instanceof FeatureConjunction) {
if (isNegated) {
return 0.0;
}
// The FeatureConjunction is handled as a leaf node in the interval algorithm.
IndexableFeatureConjunction ifc = new IndexableFeatureConjunction((FeatureConjunction) predicate);
return 1.0 / context.conjunctionOccurrences.get(ifc.id);
} else if (predicate instanceof Disjunction) {
// Minimum value of children.
return ((Disjunction) predicate).getOperands().stream().mapToDouble(child -> findMinFeature(child, isNegated, context)).min().getAsDouble();
} else if (predicate instanceof Negation) {
return findMinFeature(((Negation) predicate).getOperand(), !isNegated, context);
} else if (predicate instanceof FeatureSet) {
if (isNegated) {
return 0.0;
}
double minFeature = 1.0;
FeatureSet featureSet = (FeatureSet) predicate;
for (String value : featureSet.getValues()) {
long featureHash = Feature.createHash(featureSet.getKey(), value);
// Clever mathematics to handle scenarios where same feature is used several places in predicate tree.
minFeature = Math.min(minFeature, 1.0 / context.featureOccurrences.get(featureHash));
}
return minFeature;
} else if (predicate instanceof FeatureRange) {
if (isNegated) {
return 0.0;
}
return 1.0 / context.featureOccurrences.get(PredicateHash.hash64(((FeatureRange) predicate).getKey()));
} else {
throw new UnsupportedOperationException("Cannot handle predicate of type " + predicate.getClass().getSimpleName());
}
}
Aggregations