首页 > 代码库 > Solr4.8.0源码分析(6)之非排序查询
Solr4.8.0源码分析(6)之非排序查询
Solr4.8.0源码分析(6)之非排序查询
上篇文章简单介绍了Solr的查询流程,本文开始将详细介绍下查询的细节。查询主要分为排序查询和非排序查询,由于两者走的是两个分支,所以本文先介绍下非排序的查询。
查询的流程主要在SolrIndexSearch.getDocListC(QueryResult qr, QueryCommand cmd),顾名思义该函数对queryResultCache进行处理,并根据查询条件选择进入排序查询还是非排序查询。
1 /**
2 * getDocList version that uses+populates query and filter caches. 3 * In the event of a timeout, the cache is not populated. 4 */ 5 private void getDocListC(QueryResult qr, QueryCommand cmd) throws IOException { 6 DocListAndSet out = new DocListAndSet(); 7 qr.setDocListAndSet(out); 8 QueryResultKey key=null; 9 int maxDocRequested = cmd.getOffset() + cmd.getLen(); //当有偏移的查询产生,Solr首先会获取cmd.getOffset()+cmd.getLen()个的doc id然后 //再根据偏移量获取子集,所以maxDocRequested是实际的查询个数。 10 // check for overflow, and check for # docs in index 11 if (maxDocRequested < 0 || maxDocRequested > maxDoc()) maxDocRequested = maxDoc();// 最多的情况获取所有doc id 12 int supersetMaxDoc= maxDocRequested; 13 DocList superset = null; 14 15 int flags = cmd.getFlags(); 16 Query q = cmd.getQuery(); 17 if (q instanceof ExtendedQuery) { 18 ExtendedQuery eq = (ExtendedQuery)q; 19 if (!eq.getCache()) { 20 flags |= (NO_CHECK_QCACHE | NO_SET_QCACHE | NO_CHECK_FILTERCACHE); 21 } 22 } 23 24 25 // we can try and look up the complete query in the cache. 26 // we can‘t do that if filter!=null though (we don‘t want to 27 // do hashCode() and equals() for a big DocSet).
// 先从查询结果的缓存区查找是否出现过该条件的查询,若出现过则返回缓存的结果.关于缓存的内容将会独立写一篇文章 28 if (queryResultCache != null && cmd.getFilter()==null 29 && (flags & (NO_CHECK_QCACHE|NO_SET_QCACHE)) != ((NO_CHECK_QCACHE|NO_SET_QCACHE))) 30 { 31 // all of the current flags can be reused during warming, 32 // so set all of them on the cache key. 33 key = new QueryResultKey(q, cmd.getFilterList(), cmd.getSort(), flags); 34 if ((flags & NO_CHECK_QCACHE)==0) { 35 superset = queryResultCache.get(key); 36 37 if (superset != null) { 38 // check that the cache entry has scores recorded if we need them 39 if ((flags & GET_SCORES)==0 || superset.hasScores()) { 40 // NOTE: subset() returns null if the DocList has fewer docs than 41 // requested 42 out.docList = superset.subset(cmd.getOffset(),cmd.getLen()); //如果有缓存,就从中去除一部分子集 43 } 44 } 45 if (out.docList != null) { 46 // found the docList in the cache... now check if we need the docset too. 47 // OPT: possible future optimization - if the doclist contains all the matches, 48 // use it to make the docset instead of rerunning the query.
//获取缓存中的docSet,并传给result。 49 if (out.docSet==null && ((flags & GET_DOCSET)!=0) ) { 50 if (cmd.getFilterList()==null) { 51 out.docSet = getDocSet(cmd.getQuery()); 52 } else { 53 List<Query> newList = new ArrayList<>(cmd.getFilterList().size()+1); 54 newList.add(cmd.getQuery()); 55 newList.addAll(cmd.getFilterList()); 56 out.docSet = getDocSet(newList); 57 } 58 } 59 return; 60 } 61 } 62 63 // If we are going to generate the result, bump up to the 64 // next resultWindowSize for better caching. 65 // 修改supersetMaxDoc为queryResultWindwSize的整数倍 66 if ((flags & NO_SET_QCACHE) == 0) { 67 // handle 0 special case as well as avoid idiv in the common case. 68 if (maxDocRequested < queryResultWindowSize) { 69 supersetMaxDoc=queryResultWindowSize; 70 } else { 71 supersetMaxDoc = ((maxDocRequested -1)/queryResultWindowSize + 1)*queryResultWindowSize; 72 if (supersetMaxDoc < 0) supersetMaxDoc=maxDocRequested; 73 } 74 } else { 75 key = null; // we won‘t be caching the result 76 } 77 } 78 cmd.setSupersetMaxDoc(supersetMaxDoc); 79 80 81 // OK, so now we need to generate an answer. 82 // One way to do that would be to check if we have an unordered list 83 // of results for the base query. If so, we can apply the filters and then 84 // sort by the resulting set. This can only be used if: 85 // - the sort doesn‘t contain score 86 // - we don‘t want score returned. 87 88 // check if we should try and use the filter cache 89 boolean useFilterCache=false; 90 if ((flags & (GET_SCORES|NO_CHECK_FILTERCACHE))==0 && useFilterForSortedQuery && cmd.getSort() != null && filterCache != null) { 91 useFilterCache=true; 92 SortField[] sfields = cmd.getSort().getSort(); 93 for (SortField sf : sfields) { 94 if (sf.getType() == SortField.Type.SCORE) { 95 useFilterCache=false; 96 break; 97 } 98 } 99 }100 101 if (useFilterCache) {102 // now actually use the filter cache.103 // for large filters that match few documents, this may be104 // slower than simply re-executing the query.105 if (out.docSet == null) {106 out.docSet = getDocSet(cmd.getQuery(),cmd.getFilter());107 DocSet bigFilt = getDocSet(cmd.getFilterList());108 if (bigFilt != null) out.docSet = out.docSet.intersection(bigFilt);109 }110 // todo: there could be a sortDocSet that could take a list of111 // the filters instead of anding them first...112 // perhaps there should be a multi-docset-iterator113 sortDocSet(qr, cmd); //排序查询114 } else {115 // do it the normal way...116 if ((flags & GET_DOCSET)!=0) {117 // this currently conflates returning the docset for the base query vs118 // the base query and all filters.119 DocSet qDocSet = getDocListAndSetNC(qr,cmd);120 // cache the docSet matching the query w/o filtering121 if (qDocSet!=null && filterCache!=null && !qr.isPartialResults()) filterCache.put(cmd.getQuery(),qDocSet);122 } else {123 getDocListNC(qr,cmd); //非排序查询,这也是本文的流程。124 }125 assert null != out.docList : "docList is null";126 }127 128 if (null == cmd.getCursorMark()) {129 // Kludge...130 // we can‘t use DocSlice.subset, even though it should be an identity op131 // because it gets confused by situations where there are lots of matches, but132 // less docs in the slice then were requested, (due to the cursor)133 // so we have to short circuit the call.134 // None of which is really a problem since we can‘t use caching with135 // cursors anyway, but it still looks weird to have to special case this136 // behavior based on this condition - hence the long explanation.137 superset = out.docList; //根据offset和len截取查询结果138 out.docList = superset.subset(cmd.getOffset(),cmd.getLen()); 139 } else {140 // sanity check our cursor assumptions141 assert null == superset : "cursor: superset isn‘t null";142 assert 0 == cmd.getOffset() : "cursor: command offset mismatch";143 assert 0 == out.docList.offset() : "cursor: docList offset mismatch";144 assert cmd.getLen() >= supersetMaxDoc : "cursor: superset len mismatch: " +145 cmd.getLen() + " vs " + supersetMaxDoc;146 }147 148 // lastly, put the superset in the cache if the size is less than or equal149 // to queryResultMaxDocsCached150 if (key != null && superset.size() <= queryResultMaxDocsCached && !qr.isPartialResults()) {151 queryResultCache.put(key, superset); //如果结果的个数小于或者等于queryResultMaxDocsCached则将本次查询结果放入缓存152 }153 }
进入非排序查询分支getDocListNC(),该函数内部分直接调用Lucene的IndexSearch.Search()
1 final TopDocsCollector topCollector = buildTopDocsCollector(len, cmd); //新建TopDocsCollector对象,里面会新建(offset + len(查询条 //件的len))的HitQueue,每当获取到一个符合查询条件的doc,就会将该doc id放入HitQueue,并totalhit计数加一,这个totalhit变量也就是查询结果的数量 2 Collector collector = topCollector; 3 if (terminateEarly) { 4 collector = new EarlyTerminatingCollector(collector, cmd.len); 5 } 6 if( timeAllowed > 0 ) { 7 collector = new TimeLimitingCollector(collector, TimeLimitingCollector.getGlobalCounter(), timeAllowed);
//TimeLimitingCollector的实现原理很简单,从第一个找到符合查询条件的doc id开始计时,在达到timeAllowed之前,会想查询得到的doc id放入HitQue //ue,一旦timeAllowed到了,就会立即扔出错误,中断后续的查询。这对于我们优化查询是个重要的提示 8 } 9 if (pf.postFilter != null) {10 pf.postFilter.setLastDelegate(collector);11 collector = pf.postFilter;12 }13 try {
// 进入Lucene的IndexSearch.Search()14 super.search(query, luceneFilter, collector);15 if(collector instanceof DelegatingCollector) {16 ((DelegatingCollector)collector).finish();17 }18 }19 catch( TimeLimitingCollector.TimeExceededException x ) {20 log.warn( "Query: " + query + "; " + x.getMessage() );21 qr.setPartialResults(true);22 }23 24 totalHits = topCollector.getTotalHits(); //返回totalhit的结果25 TopDocs topDocs = topCollector.topDocs(0, len); //返回优先级队列hitqueue的doc id26 populateNextCursorMarkFromTopDocs(qr, cmd, topDocs);27 28 maxScore = totalHits>0 ? topDocs.getMaxScore() : 0.0f;29 nDocsReturned = topDocs.scoreDocs.length;30 ids = new int[nDocsReturned];31 scores = (cmd.getFlags()&GET_SCORES)!=0 ? new float[nDocsReturned] : null;32 for (int i=0; i<nDocsReturned; i++) {33 ScoreDoc scoreDoc = topDocs.scoreDocs[i];34 ids[i] = scoreDoc.doc;35 if (scores != null) scores[i] = scoreDoc.score;36 }
TimeLimitingCollector统计查询结果的方法,一旦timeAllowed到了,就会立即扔出错误,中断后续的查询
/** * Calls {@link Collector#collect(int)} on the decorated {@link Collector} * unless the allowed time has passed, in which case it throws an exception. * * @throws TimeExceededException * if the time allowed has exceeded. */ @Override public void collect(final int doc) throws IOException { final long time = clock.get(); if (timeout < time) { if (greedy) { //System.out.println(this+" greedy: before failing, collecting doc: "+(docBase + doc)+" "+(time-t0)); collector.collect(doc); } //System.out.println(this+" failing on: "+(docBase + doc)+" "+(time-t0)); throw new TimeExceededException( timeout-t0, time-t0, docBase + doc ); } //System.out.println(this+" collecting: "+(docBase + doc)+" "+(time-t0)); collector.collect(doc); }
接下来开始lucece的查询过程,
1. 首先会为每一个查询条件新建一个Weight的对象,最后将所有Weight对象放入ArrayList<Weight> weights。该过程给出每个查询条件的权重,并用于后续的评分过程。
1 public BooleanWeight(IndexSearcher searcher, boolean disableCoord) 2 throws IOException { 3 this.similarity = searcher.getSimilarity(); 4 this.disableCoord = disableCoord; 5 weights = new ArrayList<>(clauses.size()); 6 for (int i = 0 ; i < clauses.size(); i++) { 7 BooleanClause c = clauses.get(i); 8 Weight w = c.getQuery().createWeight(searcher); 9 weights.add(w);10 if (!c.isProhibited()) {11 maxCoord++;12 }13 }14 }
2. 遍历所有sgement,一个接一个的查找符合查询条件的doc id。AtomicReaderContext 是包含segment的具体信息,包括doc base,num docs,这些信息室非常有用的,在实现查询优化时候很有帮助。这里需要注意的是这个collector是TopDocsCollector类型的对象,这在上面的代码中已经赋值过了。
1 /** 2 * Lower-level search API. 3 * 4 * <p> 5 * {@link Collector#collect(int)} is called for every document. <br> 6 * 7 * <p> 8 * NOTE: this method executes the searches on all given leaves exclusively. 9 * To search across all the searchers leaves use {@link #leafContexts}.10 * 11 * @param leaves 12 * the searchers leaves to execute the searches on13 * @param weight14 * to match documents15 * @param collector16 * to receive hits17 * @throws BooleanQuery.TooManyClauses If a query would exceed 18 * {@link BooleanQuery#getMaxClauseCount()} clauses.19 */20 protected void search(List<AtomicReaderContext> leaves, Weight weight, Collector collector)21 throws IOException {22 23 // TODO: should we make this24 // threaded...? the Collector could be sync‘d?25 // always use single thread:26 for (AtomicReaderContext ctx : leaves) { // search each subreader27 try {28 collector.setNextReader(ctx);29 } catch (CollectionTerminatedException e) {30 // there is no doc of interest in this reader context31 // continue with the following leaf32 continue;33 }34 BulkScorer scorer = weight.bulkScorer(ctx, !collector.acceptsDocsOutOfOrder(), ctx.reader().getLiveDocs());35 if (scorer != null) {36 try {37 scorer.score(collector);38 } catch (CollectionTerminatedException e) {39 // collection was terminated prematurely40 // continue with the following leaf41 }42 }43 }44 }
3. Weight.bulkScorer对查询条件进行评分,Lucene的多条件查询优化还是写的很不错的。Lucece会根据每个查询条件的词频对查询条件进行排序,词频小的排在前面,词频大的排在后面。这大大优化了多条件的查询。多条件查询的优化会在下文中详细介绍。
4. 最后Lucene会使用scorer.score(collector)这个过程真正的进行查询。看下Weight的两个函数,就能明白Lucene怎么进行查询统计。
1 @Override 2 public boolean score(Collector collector, int max) throws IOException { 3 // TODO: this may be sort of weird, when we are 4 // embedded in a BooleanScorer, because we are 5 // called for every chunk of 2048 documents. But, 6 // then, scorer is a FakeScorer in that case, so any 7 // Collector doing something "interesting" in 8 // setScorer will be forced to use BS2 anyways: 9 collector.setScorer(scorer);10 if (max == DocIdSetIterator.NO_MORE_DOCS) {11 scoreAll(collector, scorer);12 return false;13 } else {14 int doc = scorer.docID();15 if (doc < 0) {16 doc = scorer.nextDoc();17 }18 return scoreRange(collector, scorer, doc, max);19 }20 }
Lucece会不停的从segment获取符合查询条件的doc,并放入collector的hitqueue里面。需要注意的是这里的collector是Collector类型,是TopDocsCollector等类的父类,所以scoreAll不仅能实现获取TopDocsCollector的doc is也能获取其他查询方式的doc id。
1 static void scoreAll(Collector collector, Scorer scorer) throws IOException {2 int doc;3 while ((doc = scorer.nextDoc()) != DocIdSetIterator.NO_MORE_DOCS) {4 collector.collect(doc);5 }6 }
进入collector.collect(doc)查看TopDocsCollector的统计doc id的方式,就跟之前说的一样。
1 @Override 2 public void collect(int doc) throws IOException { 3 float score = scorer.score(); 4 5 // This collector cannot handle these scores: 6 assert score != Float.NEGATIVE_INFINITY; 7 assert !Float.isNaN(score); 8 9 totalHits++;10 if (score <= pqTop.score) {11 // Since docs are returned in-order (i.e., increasing doc Id), a document12 // with equal score to pqTop.score cannot compete since HitQueue favors13 // documents with lower doc Ids. Therefore reject those docs too.14 return;15 }16 pqTop.doc = doc + docBase;17 pqTop.score = score;18 pqTop = pq.updateTop();19 }
总结:本章详细的介绍了非排序查询的流程,主要涉及了以下几个类QueryComponent,SolrIndexSearch,TimeLimitingCollector,TopDocsCollector,IndexSearch,BulkScore,Weight. 篇幅原因,并没有将如何从segment里面获取doc id以及多条件查询是怎么实现的,这将是下一问多条件查询中详细介绍。