LUCENE-4585: PrefixTree bugs

git-svn-id: https://svn.apache.org/repos/asf/lucene/dev/trunk@1418006 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
David Wayne Smiley 2012-12-06 17:17:53 +00:00
parent ffb1558e91
commit 4eb1c502c9
7 changed files with 61 additions and 45 deletions

View File

@ -69,6 +69,12 @@ Changes in backwards compatibility policy
previous two methods returned by calling parents(), children() or siblings()
on the returned ParallelTaxonomyArrays. (Shai Erera)
* LUCENE-4585: Spatial PrefixTree based Strategies (either TermQuery or
RecursivePrefix based) MAY want to re-index if used for point data. If a
re-index is not done, then an indexed point is ~1/2 the smallest grid cell
larger and as such is slightly more likely to match a query shape.
(David Smiley)
New Features
* LUCENE-4226: New experimental StoredFieldsFormat that compresses chunks of
@ -207,16 +213,6 @@ Bug Fixes
* LUCENE-4009: Improve TermsFilter.toString (Tim Costermans via Chris
Male, Mike McCandless)
* LUCENE-4588: Benchmark's EnwikiContentSource was discarding last wiki
document and had leaking threads in 'forever' mode. (Doron Cohen)
Changes in Runtime Behavior
* LUCENE-4586: Change default ResultMode of FacetRequest to PER_NODE_IN_TREE.
This only affects requests with depth>1. If you execute such requests and
rely on the facet results being returned flat (i.e. no hierarchy), you should
set the ResultMode to GLOBAL_FLAT. (Shai Erera, Gilad Barkai)
Optimizations
* LUCENE-2221: oal.util.BitUtil was modified to use Long.bitCount and
@ -269,9 +265,6 @@ Optimizations
Users of this API can now simply obtain an instance via DocValues#getDirectSource per thread.
(Simon Willnauer)
* LUCENE-4580: DrillDown.query variants return a ConstantScoreQuery with boost set to 0.0f
so that documents scores are not affected by running a drill-down query. (Shai Erera)
Documentation
* LUCENE-4483: Refer to BytesRef.deepCopyOf in Term's constructor that takes BytesRef.
@ -287,10 +280,6 @@ Build
RandomizedContext.contexts static map. Upgrade randomized testing
to version 2.0.2 (Mike McCandless, Dawid Weiss)
* LUCENE-4589: Upgraded benchmark module's Nekohtml dependency to version
1.9.17, removing the workaround in Lucene's HTML parser for the
Turkish locale. (Uwe Schindler)
======================= Lucene 4.0.0 =======================

View File

@ -45,6 +45,8 @@ public class PointPrefixTreeFieldCacheProvider extends ShapeFieldCacheProvider<P
@Override
protected Point readShape(BytesRef term) {
scanCell = grid.getNode(term.bytes, term.offset, term.length, scanCell);
return scanCell.isLeaf() ? scanCell.getShape().getCenter() : null;
if (scanCell.getLevel() == grid.getMaxLevels() && !scanCell.isLeaf())
return scanCell.getCenter();
return null;
}
}

View File

@ -19,7 +19,11 @@ package org.apache.lucene.spatial.prefix;
import com.spatial4j.core.shape.Shape;
import com.spatial4j.core.shape.SpatialRelation;
import org.apache.lucene.index.*;
import org.apache.lucene.index.AtomicReader;
import org.apache.lucene.index.AtomicReaderContext;
import org.apache.lucene.index.DocsEnum;
import org.apache.lucene.index.Terms;
import org.apache.lucene.index.TermsEnum;
import org.apache.lucene.search.DocIdSet;
import org.apache.lucene.search.DocIdSetIterator;
import org.apache.lucene.search.Filter;
@ -110,23 +114,28 @@ RE "scan" threshold:
while(!cells.isEmpty()) {
final Node cell = cells.removeFirst();
final BytesRef cellTerm = new BytesRef(cell.getTokenBytes());
TermsEnum.SeekStatus seekStat = termsEnum.seekCeil(cellTerm);
if (seekStat == TermsEnum.SeekStatus.END)
break;
if (seekStat == TermsEnum.SeekStatus.NOT_FOUND)
if (!termsEnum.seekExact(cellTerm, true))
continue;
if (cell.getLevel() == detailLevel || cell.isLeaf()) {
docsEnum = termsEnum.docs(acceptDocs, docsEnum, 0);
addDocs(docsEnum,bits);
} else {//any other intersection
//If the next indexed term is the leaf marker, then add all of them
assert cell.getLevel() < detailLevel; //assertions help clarify logic
assert !cell.isLeaf();
//If the next indexed term just adds a leaf marker ('+') to cell,
// then add all of those docs
BytesRef nextCellTerm = termsEnum.next();
if (nextCellTerm == null)
break;
assert StringHelper.startsWith(nextCellTerm, cellTerm);
scanCell = grid.getNode(nextCellTerm.bytes, nextCellTerm.offset, nextCellTerm.length, scanCell);
if (scanCell.isLeaf()) {
if (scanCell.getLevel() == cell.getLevel() && scanCell.isLeaf()) {
docsEnum = termsEnum.docs(acceptDocs, docsEnum, 0);
addDocs(docsEnum,bits);
termsEnum.next();//move pointer to avoid potential redundant addDocs() below
//increment pointer to avoid potential redundant addDocs() below
nextCellTerm = termsEnum.next();
if (nextCellTerm == null)
break;
}
//Decide whether to continue to divide & conquer, or whether it's time to scan through terms beneath this cell.
@ -144,8 +153,13 @@ RE "scan" threshold:
if (termLevel > detailLevel)
continue;
if (termLevel == detailLevel || scanCell.isLeaf()) {
//TODO should put more thought into implications of box vs point
Shape cShape = termLevel == grid.getMaxLevels() ? scanCell.getCenter() : scanCell.getShape();
Shape cShape;
//if this cell represents a point, use the cell center vs the box
// (points never have isLeaf())
if (termLevel == grid.getMaxLevels() && !scanCell.isLeaf())
cShape = scanCell.getCenter();
else
cShape = scanCell.getShape();
if(queryShape.relate(cShape) == SpatialRelation.DISJOINT)
continue;

View File

@ -101,11 +101,11 @@ public class GeohashPrefixTree extends SpatialPrefixTree {
class GhCell extends Node {
GhCell(String token) {
super(GeohashPrefixTree.this, token);
super(token);
}
GhCell(byte[] bytes, int off, int len) {
super(GeohashPrefixTree.this, bytes, off, len);
super(bytes, off, len);
}
@Override

View File

@ -44,11 +44,14 @@ public abstract class Node implements Comparable<Node> {
private String token;//this is the only part of equality
protected SpatialRelation shapeRel;//set in getSubCells(filter), and via setLeaf().
private SpatialPrefixTree spatialPrefixTree;
/** When set via getSubCells(filter), it is the relationship between this
* cell and the given shape filter. If set via setLeaf() (to WITHIN), it is
* meant to indicate no further sub-cells are going to be provided because
* maxLevels or a detailLevel is hit. It's always null for points.
*/
protected SpatialRelation shapeRel;
protected Node(SpatialPrefixTree spatialPrefixTree, String token) {
this.spatialPrefixTree = spatialPrefixTree;
protected Node(String token) {
this.token = token;
if (token.length() > 0 && token.charAt(token.length() - 1) == (char) LEAF_BYTE) {
this.token = token.substring(0, token.length() - 1);
@ -59,8 +62,7 @@ public abstract class Node implements Comparable<Node> {
getShape();//ensure any lazy instantiation completes to make this threadsafe
}
protected Node(SpatialPrefixTree spatialPrefixTree, byte[] bytes, int off, int len) {
this.spatialPrefixTree = spatialPrefixTree;
protected Node(byte[] bytes, int off, int len) {
this.bytes = bytes;
this.b_off = off;
this.b_len = len;
@ -78,11 +80,10 @@ public abstract class Node implements Comparable<Node> {
}
private void b_fixLeaf() {
//note that non-point shapes always have the maxLevels cell set with setLeaf
if (bytes[b_off + b_len - 1] == LEAF_BYTE) {
b_len--;
setLeaf();
} else if (getLevel() == spatialPrefixTree.getMaxLevels()) {
setLeaf();
}
}
@ -90,6 +91,10 @@ public abstract class Node implements Comparable<Node> {
return shapeRel;
}
/**
* For points, this is always false. Otherwise this is true if there are no
* further cells with this prefix for the shape (always true at maxLevels).
*/
public boolean isLeaf() {
return shapeRel == SpatialRelation.WITHIN;
}
@ -133,8 +138,14 @@ public abstract class Node implements Comparable<Node> {
//public Cell getParent();
/**
* Like {@link #getSubCells()} but with the results filtered by a shape. If that shape is a {@link com.spatial4j.core.shape.Point} then it
* must call {@link #getSubCell(com.spatial4j.core.shape.Point)};
* Like {@link #getSubCells()} but with the results filtered by a shape. If
* that shape is a {@link com.spatial4j.core.shape.Point} then it
* must call {@link #getSubCell(com.spatial4j.core.shape.Point)}.
* The returned cells should have their {@link Node#shapeRel} set to their
* relation with {@code shapeFilter} for non-point. As such,
* {@link org.apache.lucene.spatial.prefix.tree.Node#isLeaf()} should be
* accurate.
* <p/>
* Precondition: Never called when getLevel() == maxLevel.
*
* @param shapeFilter an optional filter for the returned cells.

View File

@ -228,16 +228,16 @@ public class QuadPrefixTree extends SpatialPrefixTree {
class QuadCell extends Node {
public QuadCell(String token) {
super(QuadPrefixTree.this, token);
super(token);
}
public QuadCell(String token, SpatialRelation shapeRel) {
super(QuadPrefixTree.this, token);
super(token);
this.shapeRel = shapeRel;
}
QuadCell(byte[] bytes, int off, int len) {
super(QuadPrefixTree.this, bytes, off, len);
super(bytes, off, len);
}
@Override

View File

@ -156,7 +156,7 @@ public abstract class SpatialPrefixTree {
}
final Collection<Node> subCells = node.getSubCells(shape);
if (node.getLevel() == detailLevel - 1) {
if (subCells.size() < node.getSubCellsSize()) {
if (subCells.size() < node.getSubCellsSize() || node.getLevel() == 0) {
if (inclParents)
result.add(node);
for (Node subCell : subCells) {
@ -164,7 +164,7 @@ public abstract class SpatialPrefixTree {
}
result.addAll(subCells);
} else {//a bottom level (i.e. detail level) optimization where all boxes intersect, so use parent cell.
node.setLeaf();
node.setLeaf();//the cell may not be strictly within but its close
result.add(node);
}
} else {