SOLR-1165: Use Lucene Filters in conjunction with IndexSearcher.search methods

git-svn-id: https://svn.apache.org/repos/asf/lucene/solr/trunk@777969 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Yonik Seeley 2009-05-23 17:26:54 +00:00
parent 5a39044cc0
commit ea43e4c997
3 changed files with 170 additions and 77 deletions

View File

@ -264,6 +264,9 @@ Optimizations
is ordered for easier impelemntation of skipTo, and is faster
in the general case. (yonik)
12. SOLR-1165: Use Lucene Filters and pass them down to the Lucene
search methods to filter earlier and improve performance. (yonik)
Bug Fixes
----------------------

View File

@ -885,9 +885,11 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
// do it the normal way...
cmd.setSupersetMaxDoc(supersetMaxDoc);
if ((cmd.getFlags() & GET_DOCSET)!=0) {
// this currently conflates returning the docset for the base query vs
// the base query and all filters.
DocSet qDocSet = getDocListAndSetNC(qr,cmd);
// cache the docSet matching the query w/o filtering
if (filterCache!=null && !qr.isPartialResults()) filterCache.put(cmd.getQuery(),qDocSet);
if (qDocSet!=null && filterCache!=null && !qr.isPartialResults()) filterCache.put(cmd.getQuery(),qDocSet);
} else {
getDocListNC(qr,cmd);
//Parameters: cmd.getQuery(),theFilt,cmd.getSort(),0,supersetMaxDoc,cmd.getFlags(),cmd.getTimeAllowed(),responseHeader);
@ -922,15 +924,15 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
Query query = QueryUtils.makeQueryable(cmd.getQuery());
final Filter luceneFilter = filter==null ? null : filter.getTopFilter();
// handle zero case...
if (lastDocRequested<=0) {
final DocSet filt = filter;
final float[] topscore = new float[] { Float.NEGATIVE_INFINITY };
final int[] numHits = new int[1];
HitCollector hc = new HitCollector() {
public void collect(int doc, float score) {
if (filt!=null && !filt.exists(doc)) return;
numHits[0]++;
if (score > topscore[0]) topscore[0]=score;
}
@ -939,7 +941,7 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, hc );
super.search(query, luceneFilter, hc);
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
@ -955,15 +957,12 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
// can't use TopDocs if there is a sort since it
// will do automatic score normalization.
// NOTE: this changed late in Lucene 1.9
final DocSet filt = filter;
final int[] numHits = new int[1];
final FieldSortedHitQueue hq = new FieldSortedHitQueue(reader, cmd.getSort().getSort(), len);
HitCollector hc = new HitCollector() {
private FieldDoc reusableFD;
public void collect(int doc, float score) {
if (filt!=null && !filt.exists(doc)) return;
numHits[0]++;
if (reusableFD == null)
reusableFD = new FieldDoc(doc, score);
@ -978,7 +977,7 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, hc );
super.search(query, luceneFilter, hc );
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
@ -1002,14 +1001,11 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
// No Sort specified (sort by score descending)
// This case could be done with TopDocs, but would currently require
// getting a BitSet filter from a DocSet which may be inefficient.
final DocSet filt = filter;
final ScorePriorityQueue hq = new ScorePriorityQueue(lastDocRequested);
final int[] numHits = new int[1];
HitCollector hc = new HitCollector() {
private ScoreDoc reusableSD;
public void collect(int doc, float score) {
if (filt!=null && !filt.exists(doc)) return;
// TODO: if docs are always delivered in order, we could use "score>minScore"
// instead of "score>=minScore" and avoid tiebreaking scores
// in the priority queue.
@ -1033,7 +1029,7 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, hc );
super.search(query, luceneFilter, hc );
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
@ -1110,10 +1106,10 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
}
// the DocSet returned is for the query only, without any filtering... that way it may
// be cached if desired.
private DocSet getDocListAndSetNC(QueryResult qr,QueryCommand cmd) throws IOException {
///////////////////// NEW
int len = cmd.getSupersetMaxDoc();
DocSet filter = cmd.getFilter()!=null ? cmd.getFilter() : getDocSet(cmd.getFilterList());
int last = len;
@ -1124,22 +1120,13 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
float maxScore;
int[] ids;
float[] scores;
final DocSetHitCollector setHC = new DocSetHitCollector(maxDoc()>>6, maxDoc());
final HitCollector collector = ( cmd.getTimeAllowed() > 0 ) ? new TimeLimitedCollector( setHC, cmd.getTimeAllowed() ) : setHC;
final DocSetHitCollector collector = new DocSetHitCollector(maxDoc()>>6, maxDoc());
Query query = QueryUtils.makeQueryable(cmd.getQuery());
final long timeAllowed = cmd.getTimeAllowed();
// TODO: perhaps unify getDocListAndSetNC and getDocListNC without imposing a significant performance hit
// Comment: gathering the set before the filter is applied allows one to cache
// the resulting DocSet under the query. The drawback is that it requires an
// extra intersection with the filter at the end. This will be a net win
// for expensive queries.
// Q: what if the final intersection results in a small set from two large
// sets... it won't be a HashDocSet or other small set. One way around
// this would be to collect the resulting set as we go (the filter is
// checked anyway).
final Filter luceneFilter = filter==null ? null : filter.getTopFilter();
// handle zero case...
if (lastDocRequested<=0) {
@ -1147,22 +1134,26 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
final float[] topscore = new float[] { Float.NEGATIVE_INFINITY };
final int[] numHits = new int[1];
try {
super.search(query, new HitCollector() {
HitCollector hc = new HitCollector() {
public void collect(int doc, float score) {
collector.collect(doc,score);
if (filt!=null && !filt.exists(doc)) return;
collector.collect(doc, score);
numHits[0]++;
if (score > topscore[0]) topscore[0]=score;
}
}
);
};
if( timeAllowed > 0 ) {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, luceneFilter, hc);
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
qr.setPartialResults(true);
}
nDocsReturned=0;
ids = new int[nDocsReturned];
scores = new float[nDocsReturned];
@ -1173,16 +1164,13 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
// will do automatic score normalization.
// NOTE: this changed late in Lucene 1.9
final DocSet filt = filter;
final int[] numHits = new int[1];
final FieldSortedHitQueue hq = new FieldSortedHitQueue(reader, cmd.getSort().getSort(), len);
try {
super.search(query, new HitCollector() {
HitCollector hc = new HitCollector() {
private FieldDoc reusableFD;
public void collect(int doc, float score) {
collector.collect(doc,score);
if (filt!=null && !filt.exists(doc)) return;
collector.collect(doc, score);
numHits[0]++;
if (reusableFD == null)
reusableFD = new FieldDoc(doc, score);
@ -1192,14 +1180,20 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
}
reusableFD = (FieldDoc) hq.insertWithOverflow(reusableFD);
}
}
);
};
if( timeAllowed > 0 ) {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, luceneFilter, hc);
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
qr.setPartialResults(true);
}
totalHits = numHits[0];
maxScore = totalHits>0 ? hq.getMaxScore() : 0.0f;
@ -1218,33 +1212,37 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
// This case could be done with TopDocs, but would currently require
// getting a BitSet filter from a DocSet which may be inefficient.
final DocSet filt = filter;
final ScorePriorityQueue hq = new ScorePriorityQueue(lastDocRequested);
final int[] numHits = new int[1];
try {
super.search(query, new HitCollector() {
HitCollector hc = new HitCollector() {
private ScoreDoc reusableSD;
public void collect(int doc, float score) {
collector.collect(doc,score);
if (filt!=null && !filt.exists(doc)) return;
// if docs are always delivered in order, we could use "score>minScore"
// but might BooleanScorer14 might still be used and deliver docs out-of-order?
int nhits = numHits[0]++;
if (reusableSD == null) {
reusableSD = new ScoreDoc(doc, score);
} else if (nhits < lastDocRequested || score >= reusableSD.score) {
// reusableSD holds the last "rejected" entry, so, if
// this new score is not better than that, there's no
// need to try inserting it
reusableSD.doc = doc;
reusableSD.score = score;
} else {
return;
}
reusableSD = (ScoreDoc) hq.insertWithOverflow(reusableSD);
collector.collect(doc, score);
// if docs are always delivered in order, we could use "score>minScore"
// but might BooleanScorer14 might still be used and deliver docs out-of-order?
int nhits = numHits[0]++;
if (reusableSD == null) {
reusableSD = new ScoreDoc(doc, score);
} else if (nhits < lastDocRequested || score >= reusableSD.score) {
// reusableSD holds the last "rejected" entry, so, if
// this new score is not better than that, there's no
// need to try inserting it
reusableSD.doc = doc;
reusableSD.score = score;
} else {
return;
}
}
);
reusableSD = (ScoreDoc) hq.insertWithOverflow(reusableSD);
}
};
if( timeAllowed > 0 ) {
hc = new TimeLimitedCollector( hc, timeAllowed );
}
try {
super.search(query, luceneFilter, hc);
}
catch( TimeLimitedCollector.TimeExceededException x ) {
log.warn( "Query: " + query + "; " + x.getMessage() );
@ -1269,11 +1267,14 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
if (sliceLen < 0) sliceLen=0;
qr.setDocList(new DocSlice(0,sliceLen,ids,scores,totalHits,maxScore));
DocSet qDocSet = setHC.getDocSet();
qr.setDocSet(filter==null ? qDocSet : qDocSet.intersection(filter));
return qDocSet;
}
// TODO: if we collect results before the filter, we just need to intersect with
// that filter to generate the DocSet for qr.setDocSet()
qr.setDocSet(collector.getDocSet());
// TODO: currently we don't generate the DocSet for the base query.
// But the QueryDocSet == CompleteDocSet if filter==null.
return filter==null ? qr.getDocSet() : null;
}
/**
@ -1701,7 +1702,6 @@ public class SolrIndexSearcher extends IndexSearcher implements SolrInfoMBean {
private int supersetMaxDoc;
private int flags;
private long timeAllowed = -1;
private boolean needDocSet;
public Query getQuery() { return query; }
public QueryCommand setQuery(Query query) {

View File

@ -18,10 +18,7 @@
package org.apache.solr.search;
import org.apache.lucene.index.Term;
import org.apache.lucene.search.TermQuery;
import org.apache.lucene.search.Query;
import org.apache.lucene.search.BooleanQuery;
import org.apache.lucene.search.BooleanClause;
import org.apache.lucene.search.*;
import org.apache.solr.util.AbstractSolrTestCase;
import org.apache.solr.request.SolrQueryRequest;
import org.apache.solr.request.SolrQueryResponse;
@ -31,7 +28,7 @@ import org.apache.solr.update.AddUpdateCommand;
import org.apache.solr.common.params.UpdateParams;
import org.apache.solr.common.SolrInputDocument;
import java.util.Random;
import java.util.*;
import java.io.IOException;
/**
@ -71,16 +68,50 @@ public class TestSearchPerf extends AbstractSolrTestCase {
}
void createIndex2(int nDocs) throws IOException {
// Skip encoding for updating the index
void createIndex2(int nDocs, String... fields) throws IOException {
Set<String> fieldSet = new HashSet<String>(Arrays.asList(fields));
SolrQueryRequest req = lrf.makeRequest();
SolrQueryResponse rsp = new SolrQueryResponse();
UpdateRequestProcessorChain processorChain = req.getCore().getUpdateProcessingChain(null);
UpdateRequestProcessor processor = processorChain.createProcessor(req, rsp);
boolean foomany_s = fieldSet.contains("foomany_s");
boolean foo1_s = fieldSet.contains("foo1_s");
boolean foo2_s = fieldSet.contains("foo2_s");
boolean foo4_s = fieldSet.contains("foo4_s");
boolean foo8_s = fieldSet.contains("foo8_s");
boolean t10_100_ws = fieldSet.contains("t10_100_ws");
for (int i=0; i<nDocs; i++) {
SolrInputDocument doc = new SolrInputDocument();
doc.addField("id",Float.toString(i));
doc.addField("foomany_s",t(r.nextInt(nDocs*10)));
if (foomany_s) {
doc.addField("foomany_s",t(r.nextInt(nDocs*10)));
}
if (foo1_s) {
doc.addField("foo1_s",t(0));
}
if (foo2_s) {
doc.addField("foo2_s",r.nextInt(2));
}
if (foo4_s) {
doc.addField("foo4_s",r.nextInt(4));
}
if (foo8_s) {
doc.addField("foo8_s",r.nextInt(8));
}
if (t10_100_ws) {
StringBuilder sb = new StringBuilder(9*100);
for (int j=0; j<100; j++) {
sb.append(' ');
sb.append(t(r.nextInt(10)));
}
doc.addField("t10_100_ws", sb.toString());
}
AddUpdateCommand cmd = new AddUpdateCommand();
cmd.solrDoc = doc;
processor.processAdd(cmd);
@ -113,6 +144,34 @@ public class TestSearchPerf extends AbstractSolrTestCase {
System.out.println("ret="+ret+ " time="+(end-start)+" throughput="+iter*1000/(end-start+1));
req.close();
assertTrue(ret>0); // make sure we did some work
return ret;
}
int doListGen(int iter, Query q, List<Query> filt, boolean cacheQuery, boolean cacheFilt) throws Exception {
SolrQueryRequest req = lrf.makeRequest();
SolrIndexSearcher searcher = req.getSearcher();
long start = System.currentTimeMillis();
// These aren't public in SolrIndexSearcher
int NO_CHECK_QCACHE = 0x80000000;
int GET_DOCSET = 0x40000000;
int NO_CHECK_FILTERCACHE = 0x20000000;
int GET_SCORES = 0x01;
int ret = 0;
for (int i=0; i<iter; i++) {
DocList l = searcher.getDocList(q, filt, (Sort)null, 0, 10, (cacheQuery?0:NO_CHECK_QCACHE)|(cacheFilt?0:NO_CHECK_FILTERCACHE) );
ret += l.matches();
}
long end = System.currentTimeMillis();
System.out.println("ret="+ret+ " time="+(end-start)+" throughput="+iter*1000/(end-start+1));
req.close();
assertTrue(ret>0); // make sure we did some work
return ret;
}
@ -132,7 +191,7 @@ public class TestSearchPerf extends AbstractSolrTestCase {
/** test range query performance */
public void XtestRangePerformance() throws Exception {
int indexSize=199999;
int indexSize=1999;
float fractionCovered=1.0f;
String l=t(0);
@ -146,7 +205,7 @@ public class TestSearchPerf extends AbstractSolrTestCase {
Query frange = parser2.parse();
req.close();
createIndex2(indexSize);
createIndex2(indexSize,"foomany_s");
doSetGen(1, range);
doSetGen(1, frange); // load field cache
@ -155,4 +214,35 @@ public class TestSearchPerf extends AbstractSolrTestCase {
doSetGen(10000, frange);
}
/** test range query performance */
public void testFilteringPerformance() throws Exception {
int indexSize=19999;
float fractionCovered=.1f;
String l=t(0);
String u=t((int)(indexSize*10*fractionCovered));
SolrQueryRequest req = lrf.makeRequest();
QParser parser = QParser.getParser("foomany_s:[" + l + " TO " + u + "]", null, req);
Query rangeQ = parser.parse();
List<Query> filters = new ArrayList<Query>();
filters.add(rangeQ);
req.close();
parser = QParser.getParser("{!dismax qf=t10_100_ws pf=t10_100_ws ps=20}"+ t(0) + ' ' + t(1) + ' ' + t(2), null, req);
Query q= parser.parse();
// SolrIndexSearcher searcher = req.getSearcher();
// DocSet range = searcher.getDocSet(rangeQ, null);
createIndex2(indexSize, "foomany_s", "t10_100_ws");
// doListGen(100, q, filters, false, true);
doListGen(500, q, filters, false, true);
req.close();
}
}