| High-performance single-document main memory Apache Lucene fulltext search index.
Overview
This class is a replacement/substitute for a large subset of
org.apache.lucene.store.RAMDirectory functionality. It is designed to
enable maximum efficiency for on-the-fly matchmaking combining structured and
fuzzy fulltext search in realtime streaming applications such as Nux XQuery based XML
message queues, publish-subscribe systems for Blogs/newsfeeds, text chat, data acquisition and
distribution systems, application level routers, firewalls, classifiers, etc.
Rather than targetting fulltext search of infrequent queries over huge persistent
data archives (historic search), this class targets fulltext search of huge
numbers of queries over comparatively small transient realtime data (prospective
search).
For example as in
float score = search(String text, Query query)
Each instance can hold at most one Lucene "document", with a document containing
zero or more "fields", each field having a name and a fulltext value. The
fulltext value is tokenized (split and transformed) into zero or more index terms
(aka words) on addField(), according to the policy implemented by an
Analyzer. For example, Lucene analyzers can split on whitespace, normalize to lower case
for case insensitivity, ignore common terms with little discriminatory value such as "he", "in", "and" (stop
words), reduce the terms to their natural linguistic root form such as "fishing"
being reduced to "fish" (stemming), resolve synonyms/inflexions/thesauri
(upon indexing and/or querying), etc. For details, see
Lucene Analyzer Intro.
Arbitrary Lucene queries can be run against this class - see Lucene Query Syntax
as well as Query Parser Rules.
Note that a Lucene query selects on the field names and associated (indexed)
tokenized terms, not on the original fulltext(s) - the latter are not stored
but rather thrown away immediately after tokenization.
For some interesting background information on search technology, see Bob Wyman's
Prospective Search,
Jim Gray's
A Call to Arms - Custom subscriptions, and Tim Bray's
On Search, the Series.
Example Usage
Analyzer analyzer = PatternAnalyzer.DEFAULT_ANALYZER;
//Analyzer analyzer = new SimpleAnalyzer();
MemoryIndex index = new MemoryIndex();
index.addField("content", "Readings about Salmons and other select Alaska fishing Manuals", analyzer);
index.addField("author", "Tales of James", analyzer);
QueryParser parser = new QueryParser("content", analyzer);
float score = index.search(parser.parse("+author:james +salmon~ +fish* manual~"));
if (score > 0.0f) {
System.out.println("it's a match");
} else {
System.out.println("no match found");
}
System.out.println("indexData=" + index.toString());
Example XQuery Usage
(: An XQuery that finds all books authored by James that have something to do with "salmon fishing manuals", sorted by relevance :)
declare namespace lucene = "java:nux.xom.pool.FullTextUtil";
declare variable $query := "+salmon~ +fish* manual~"; (: any arbitrary Lucene query can go here :)
for $book in /books/book[author="James" and lucene:match(abstract, $query) > 0.0]
let $score := lucene:match($book/abstract, $query)
order by $score descending
return $book
No thread safety guarantees
An instance can be queried multiple times with the same or different queries,
but an instance is not thread-safe. If desired use idioms such as:
MemoryIndex index = ...
synchronized (index) {
// read and/or write index (i.e. add fields and/or query)
}
Performance Notes
Internally there's a new data structure geared towards efficient indexing
and searching, plus the necessary support code to seamlessly plug into the Lucene
framework.
This class performs very well for very small texts (e.g. 10 chars)
as well as for large texts (e.g. 10 MB) and everything in between.
Typically, it is about 10-100 times faster than RAMDirectory.
Note that RAMDirectory has particularly
large efficiency overheads for small to medium sized texts, both in time and space.
Indexing a field with N tokens takes O(N) in the best case, and O(N logN) in the worst
case. Memory consumption is probably larger than for RAMDirectory.
Example throughput of many simple term queries over a single MemoryIndex:
~500000 queries/sec on a MacBook Pro, jdk 1.5.0_06, server VM.
As always, your mileage may vary.
If you're curious about
the whereabouts of bottlenecks, run java 1.5 with the non-perturbing '-server
-agentlib:hprof=cpu=samples,depth=10' flags, then study the trace log and
correlate its hotspot trailer with its call stack headers (see
hprof tracing ).
author:
whoschek.AT.lbl.DOT.gov |
