Good point.
By the way, by now I have a version against 1.4.3 that is 10-100 times faster (i.e. 30000 - 200000 index+query steps/sec) than the simplistic RAMDirectory approach, depending on the nature of the input data and query. From some preliminary testing it returns exactly what RAMDirectory returns.
I'll do some cleanup and documentation and then post this to the list for review RSN.
As an aside, is there any work going on to potentially support prefix (and infix) wild card queries ala "*fish"?
Wolfgang.
On Apr 20, 2005, at 6:10 AM, Vanlerberghe, Luc wrote:
One reason to choose the 'simplistic IndexReader' approach to this problem over regex's is that the result should be 'bug-compatible' with a standard search over all documents.
Differences between the two systems would be difficult to explain to an end-user (let alone for the developer to debug and find the reason in the first place!)
Luc
-----Original Message----- From: Erik Hatcher [mailto:[EMAIL PROTECTED] Sent: Saturday, April 16, 2005 2:09 AM To: java-dev@lucene.apache.org Subject: Re: [Performance] Streaming main memory indexing of single strings
On Apr 15, 2005, at 6:15 PM, Wolfgang Hoschek wrote:Cool! For my use case it would need to be able to handle arbitrary queries (previously parsed from a general lucene query string). Something like:
float match(String Text, Query query)
it's fine with me if it also works for
float[] match(String[] texts, Query query) or float(Document doc, Query query)
but that isn't required by the use case.
My implementation is nearly that. The score is available as
hits.score(0). You would also need an analyzer, I presume, passed to
your proposed match() method if you want the text broken into terms.
My current implementation is passed a String[] where each item is
considered a term for the document. match() would also need a field
name to be fully accurate - since the analyzer needs a field name and
terms used for searching need a field name. The Query may contain terms
for any number of fields - how should that be handled? Should only a
single field name be passed in and any terms request for other fields be
ignored? Or should this utility morph to assume any words in the text
is in any field being asked of it?
As for Doug's devil advocate questions - I really don't know what I'd
use it for personally (other than the "match this single string against
a bunch of queries"), I just thought it was clever that it could be
done. Clever regex's could come close, but it'd be a lot more effort
than reusing good ol' QueryParser and this simplistic IndexReader, along
with an Analyzer.
Erik
Wolfgang.
I am intrigued by this and decided to mock a quick and dirty example of such an IndexReader. After a little trial-and-error I got it working at least for TermQuery and WildcardQuery. I've pasted my code below as an example, but there is much room for improvement, especially in terms of performance and also in keeping track of term frequency, and also it would be nicer if it handled the analysis internally.
I think something like this would make a handy addition to our contrib area at least. I'd be happy to receive improvements to this and then add it to a contrib subproject.
Perhaps this would be a handy way to handle situations where users have queries saved in a system and need to be alerted whenever a new document arrives matching the saved queries?
Erik
-----Original Message----- From: Wolfgang Hoschek [mailto:[EMAIL PROTECTED] Sent: Thursday, April 14, 2005 4:04 PM To: java-dev@lucene.apache.org Subject: Re: [Performance] Streaming main memory indexing of single strings
This seems to be a promising avenue worth exploring. My gutfeeling is that this could easily be 10-100 times faster.
The drawback is that it requires a fair amount of understanding of intricate Lucene internals, pulling those pieces together and adapting them as required for the seemingly simple "float match(String text, Query query)".
I might give it a shot but I'm not sure I'll be able to pull this off! Is there any similar code I could look at as a starting point?
Wolfgang.
On Apr 14, 2005, at 1:13 PM, Robert Engels wrote:
I think you are not approaching this the correct way.
Pseudo code:
Subclass IndexReader.
Get tokens from String 'document' using Lucene analyzers.
Build simple hash-map based data structures using tokens for terms,
and term positions.
reimplement termDocs() and termPositions() to use the structures from above.
run searches.
start again with next document.
-----Original Message----- From: Wolfgang Hoschek [mailto:[EMAIL PROTECTED] Sent: Thursday, April 14, 2005 2:56 PM To: java-dev@lucene.apache.org Subject: Re: [Performance] Streaming main memory indexing of single
strings
Otis, this might be a misunderstanding.
- I'm not calling optimize(). That piece is commented out you if look again at the code. - The *streaming* use case requires that for each query I add one (and only one) document (aka string) to an empty index:
repeat N times (where N is millions or billions): add a single string (aka document) to an empty index query the index drop index (or delete it's document)
with the following API being called N times: float match(String text, Query query)
So there's no possibility of adding many documents and thereafter running the query. This in turn seems to mean that the IndexWriter can't be kept open - unless I manually delete each document after each query to repeatedly reuse the RAMDirectory, which I've also tried before without any significant performance gain - deletion seems to have substantial overhead in itself. Perhaps it would be better if there were a Directory.deleteAllDocuments() or similar. Did you have some other approach in mind?
As I said, Lucene's design doesn't seem to fit this streaming use case pattern well. In *this* scenario one could easily do without any locking, and without byte level organization in RAMDirectory and RAMFile, etc because a single small string isn't a large persistent multi-document index.
For some background, here's a small example for the kind of XQuery functionality Nux/Lucene integration enables:
(: An XQuery that finds all books authored by James that have something to do with "fish", sorted by relevance :) declare namespace lucene = "java:nux.xom.xquery.XQueryUtil"; declare variable $query := "fish*~";
for $book in /books/book[author="James" and lucene:match(string(.), $query) > 0.0] let $score := lucene:match(string($book), $query) order by $score descending return (<score>{$score}</score>, $book)
More interestingly one can use this for classifying and routing XML
messages based on rules (i.e. queries) inspecting their content...
Any other clues about potential improvements would be greatly appreciated.
Wolfgang.
On Apr 13, 2005, at 10:09 PM, Otis Gospodnetic wrote:
It looks like you are calling that IndexWriter code in some loops,
opening it and closing it in every iteration of the loop and also calling optimize. All of those things could be improved. Keep your IndexWriter open, don't close it, and optimize the index
only once you are done adding documents to it.
See the highlights and the snipets in the first hit: http://www.lucenebook.com/search?query=when+to+optimize
Otis
--- Wolfgang Hoschek <[EMAIL PROTECTED]> wrote:
Hi,
I'm wondering if anyone could let me know how to improve Lucene performance for "streaming main memory indexing of single strings". This would help to effectively integrate Lucene with the Nux XQuery engine.
Below is a small microbenchmark simulating STREAMING XQuery fulltext search as typical for XML network routers, message queuing system, P2P networks, etc. In this on-the-fly main memory
indexing scenario, each
individual string is immediately matched as soon as it becomes available without any persistance involved. This usage scenario and corresponding performance profile is quite different in comparison to
fulltext search over persistent (read-mostly) indexes.
The benchmark runs at some 3000 lucene queries/sec (lucene-1.4.3)
which is unfortunate news considering the XQuery engine can easily walk hundreds of thousands of XML nodes per second. Ideally I'd like to run at some 100000 queries/sec. Runnning this
tokenStream(String fieldName, Readerthrough the JDK 1.5 profiler it seems that most time is spent in and below the following calls:
writer = new IndexWriter(dir, analyzer, true); writer.addDocument(...); writer.close();
I tried quite a few variants of the benchmark with various options, unfortunately with little or no effect. Lucene just does not seem to designed to do this sort of "transient single string index" thing. All code paths related to opening, closing, reading, writing, querying and object creation seem to be designed for large persistent indexes.
Any advice on what I'm missing or what could be done about it would be greatly appreciated.
Wolfgang.
P.S. the benchmark code is attached as a file below:
package nux.xom.pool;
import java.io.IOException; //import java.io.Reader;
import org.apache.lucene.analysis.Analyzer; //import org.apache.lucene.analysis.LowerCaseTokenizer; //import org.apache.lucene.analysis.PorterStemFilter; //import org.apache.lucene.analysis.SimpleAnalyzer; //import org.apache.lucene.analysis.TokenStream; import org.apache.lucene.analysis.standard.StandardAnalyzer; import org.apache.lucene.document.Document; import org.apache.lucene.document.Field; //import org.apache.lucene.index.IndexReader; import org.apache.lucene.index.IndexWriter; import org.apache.lucene.queryParser.ParseException; import org.apache.lucene.queryParser.QueryParser; import org.apache.lucene.search.Hits; import org.apache.lucene.search.IndexSearcher; import org.apache.lucene.search.Query; import org.apache.lucene.search.Searcher; import org.apache.lucene.store.Directory; import org.apache.lucene.store.RAMDirectory;
public final class LuceneMatcher { // TODO: make non-public
private final Analyzer analyzer; // private final Directory dir = new RAMDirectory();
public LuceneMatcher() { this(new StandardAnalyzer()); // this(new SimpleAnalyzer()); // this(new StopAnalyzer()); // this(new Analyzer() { // public final TokenStreamLowerCaseTokenizer(reader));reader) { // return new PorterStemFilter(newmust not be// } // }); }
public LuceneMatcher(Analyzer analyzer) { if (analyzer == null) throw new IllegalArgumentException("analyzernull"); this.analyzer = analyzer; }
public Query parseQuery(String expression) throws ParseException
analyzer);{ QueryParser parser = new QueryParser("content",// parser.setPhraseSlop(0); return parser.parse(expression); }
/** * Returns the relevance score by matching the given index against the given * Lucene query expression. The index must not contain more than
: 0.0f;one Lucene * "document" (aka string to be searched). */ public float match(Directory index, Query query) { Searcher searcher = null; try { searcher = new IndexSearcher(index); Hits hits = searcher.search(query); float score = hits.length() > 0 ? hits.score(0)(RAMDirectory)return score; } catch (IOException e) { // should never happen(RAMDirectory)throw new RuntimeException(e); } finally { try { if (searcher != null) searcher.close(); } catch (IOException e) { // should never happen(RAMDirectory)throw new RuntimeException(e); } } }
// public float match(String text, Query query) { // return match(createIndex(text), query); // }
public Directory createIndex(String text) { Directory dir = new RAMDirectory(); IndexWriter writer = null; try { writer = new IndexWriter(dir, analyzer, true); // writer.setUseCompoundFile(false); // writer.mergeFactor = 2; // writer.minMergeDocs = 1; // writer.maxMergeDocs = 1;
writer.addDocument(createDocument(text)); // writer.optimize(); return dir; } catch (IOException e) { // should never happen(RAMDirectory)throw new RuntimeException(e); } finally { try { if (writer != null) writer.close(); } catch (IOException e) { // should never happenthrow new RuntimeException(e); } } }
private Document createDocument(String content) { Document doc = new Document(); doc.add(Field.UnStored("content", content)); // doc.add(Field.Text("x", content)); return doc; }
/** * Lucene microbenchmark simulating STREAMING XQuery fulltext search as * typical for XML network routers, message queuing system, P2P networks, * etc. In this on-the-fly main memory indexing scenario, each individual * string is immediately matched as soon as it becomes available
Integer.parseInt(args[k]);without any * persistance involved. This usage scenario and corresponding performance * profile is quite different in comparison to fulltext search over * persistent (read-mostly) indexes. * * Example XPath: count(/table/row[lucene:match(string(./firstname), * "James") > 0.0]) */ public static void main(String[] args) throws Exception { int k = -1; int runs = 5; if (args.length > ++k) runs = Integer.parseInt(args[k]);
int nodes = 10000; if (args.length > ++k) nodes =reused N
String content = "James is out in the woods"; if (args.length > ++k) content = args[k];
String expression = "James"; if (args.length > ++k) expression = args[k];
LuceneMatcher matcher = new LuceneMatcher(); Query query = matcher.parseQuery(expression); // to bei, expression) > 0.0f) {times
for (int r = 0; r < runs; r++) { long start = System.currentTimeMillis(); int matches = 0;
for (int i = 0; i < nodes; i++) { // if (LuceneUtil.match(content +(matcher.match(matcher.createIndex(content + i), query) >if1000.0f));0.0f) { matches++; } }
long end = System.currentTimeMillis(); System.out.println("matches=" + matches); System.out.println("secs=" + ((end-start) /((end-start) /System.out.println("queries/sec=" + (nodes /1000.0f))); System.out.println(); } } }
public class StringIndexReader extends IndexReader { private List terms; public StringIndexReader(String strings[]) { super(null); terms = Arrays.asList(strings); Collections.sort(terms); }
public TermFreqVector[] getTermFreqVectors(int docNumber) throws IOException { System.out.println("StringIndexReader.getTermFreqVectors"); return new TermFreqVector[0]; }
public TermFreqVector getTermFreqVector(int docNumber, String field) throws IOException { System.out.println("StringIndexReader.getTermFreqVector"); return null; }
public int numDocs() { System.out.println("StringIndexReader.numDocs"); return 1; }
public int maxDoc() { System.out.println("StringIndexReader.maxDoc"); return 1; }
public Document document(int n) throws IOException { System.out.println("StringIndexReader.document"); return null; }
public boolean isDeleted(int n) { System.out.println("StringIndexReader.isDeleted"); return false; }
public boolean hasDeletions() { System.out.println("StringIndexReader.hasDeletions"); return false; }
public byte[] norms(String field) throws IOException { // TODO: what value to use for this? System.out.println("StringIndexReader.norms: " + field); return new byte[] { 1 }; }
public void norms(String field, byte[] bytes, int offset) throws IOException { System.out.println("StringIndexReader.norms: " + field + "*"); }
protected void doSetNorm(int doc, String field, byte value) throws IOException { System.out.println("StringIndexReader.doSetNorm");
}
public TermEnum terms() throws IOException { System.out.println("StringIndexReader.terms"); return terms(null); }
public TermEnum terms(final Term term) throws IOException { System.out.println("StringIndexReader.terms: " + term);
TermEnum termEnum = new TermEnum() { private String currentTerm; private Iterator iter;
public boolean next() { System.out.println("TermEnum.next"); if (iter.hasNext()) currentTerm = (String) iter.next(); return iter.hasNext(); }
public Term term() { if (iter == null) { iter = terms.iterator(); while (next()) { if (currentTerm.startsWith(term.text())) break; } } System.out.println("TermEnum.term: " + currentTerm); return new Term(term.field(), currentTerm); }
public int docFreq() { System.out.println("TermEnum.docFreq"); return 1; }
public void close() { System.out.println("TermEnum.close"); } }; return termEnum; }
public int docFreq(Term term) throws IOException { System.out.println("StringIndexReader.docFreq: " + term); return terms.contains(term.text()) ? 1 : 0; }
public TermDocs termDocs() throws IOException { System.out.println("StringIndexReader.termDocs");
TermDocs td = new TermDocs() { private boolean done = false; String currentTerm;
public void seek(Term term) { System.out.println(".seek: " + term); currentTerm = term.text(); done = false; }
public void seek(TermEnum termEnum) { seek(termEnum.term()); }
public int doc() { System.out.println(".doc"); return 0; }
public int freq() { System.out.println(".freq"); return 1; }
public boolean next() { System.out.println(".next"); return false; }
public int read(int[] docs, int[] freqs) { System.out.println(".read: " + docs.length);
if (done) return 0;
done = true; docs[0] = 0; freqs[0] = freq(); return 1; }
public boolean skipTo(int target) { System.out.println(".skipTo"); return false; }
public void close() { System.out.println(".close");
} }; return td; }
public TermPositions termPositions() throws IOException { System.out.println("StringIndexReader.termPositions"); return null; }
protected void doDelete(int docNum) throws IOException { System.out.println("StringIndexReader.doDelete");
}
protected void doUndeleteAll() throws IOException { System.out.println("StringIndexReader.doUndeleteAll");
}
protected void doCommit() throws IOException { System.out.println("StringIndexReader.doCommit");
}
protected void doClose() throws IOException { System.out.println("StringIndexReader.doClose");
}
public Collection getFieldNames() throws IOException { System.out.println("StringIndexReader.getFieldNames"); return null; }
public Collection getFieldNames(boolean indexed) throws IOException
{ System.out.println("StringIndexReader.getFieldNames"); return null; }
public Collection getIndexedFieldNames(Field.TermVector tvSpec) { System.out.println("StringIndexReader.getIndexedFieldNames"); return null; }
public Collection getFieldNames(FieldOption fldOption) { System.out.println("StringIndexReader.getFieldNames"); return null; }
public static void main(String[] args) { IndexReader reader = new StringIndexReader(new String[] {"foo", "bar", "baz"}); IndexSearcher searcher = new IndexSearcher(reader);
Hits hits = null; try { hits = searcher.search(new WildcardQuery(new Term("field","ba*"))); } catch (IOException e) { e.printStackTrace(); } System.out.println("found " + hits.length()); } }
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
----------------------------------------------------------------------- Wolfgang Hoschek | email: [EMAIL PROTECTED] Distributed Systems Department | phone: (415)-533-7610 Berkeley Laboratory | http://dsd.lbl.gov/~hoschek/ -----------------------------------------------------------------------
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
