Skip to content

Speed up loading dense singleton keyword fields #132994

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
martijnvg merged 6 commits into from
Aug 16, 2025
Merged

Speed up loading dense singleton keyword fields #132994

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
8137db1
Speed up loading dense singleton keyword fields
martijnvg Aug 15, 2025
a84961e
SingletonOrdinalsBuilder unit testing
martijnvg Aug 15, 2025
f74f9ec
Fallback to tryRead(...) if look ahead finds multiple ordinals
martijnvg Aug 15, 2025
7613265
Added BaseSortedDocValues#tryReadAHead(...) to improve testing of the…
martijnvg Aug 15, 2025
8e57dd0
off course testAllNull() can't handle deletes 🤦‍♂️...
martijnvg Aug 16, 2025
067addd
Merge remote-tracking branch 'es/main' into speedup_loading_dense_key…
martijnvg Aug 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
149 changes: 121 additions & 28 deletions ...er/src/main/java/org/elasticsearch/index/codec/tsdb/es819/ES819TSDBDocValuesProducer.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -345,6 +345,10 @@ public SortedDocValues getSorted(FieldInfo field) throws IOException {
}

private SortedDocValues getSorted(SortedEntry entry, boolean valuesSorted) throws IOException {
if (entry.ordsEntry.docsWithFieldOffset == -2) {
return DocValues.emptySorted();
}

final NumericDocValues ords = getNumeric(entry.ordsEntry, entry.termsDictEntry.termsDictSize);
return new BaseSortedDocValues(entry) {

Expand Down Expand Up @@ -380,7 +384,25 @@ public long cost() {

@Override
public BlockLoader.Block tryRead(BlockLoader.BlockFactory factory, BlockLoader.Docs docs, int offset) throws IOException {
if (valuesSorted && ords instanceof BaseDenseNumericValues denseOrds) {
if (ords instanceof BaseDenseNumericValues denseOrds) {
var block = tryReadAHead(factory, docs, offset);
if (block != null) {
return block;
}
// Falling back to tryRead(...) is safe here, given that current block index wasn't altered by looking ahead.
try (var builder = factory.singletonOrdinalsBuilder(this, docs.count() - offset, true)) {
BlockLoader.SingletonLongBuilder delegate = new SingletonLongToSingletonOrdinalDelegate(builder);
Copy link
Member Author

@martijnvg martijnvg Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any subclasses of BlockLoader.Block aren't accessible from server. So instead of consuming the block that denseOrds would produce, I have a delegate builder that translates the long values to int values (given that singleton ordinals works with int[]).

var result = denseOrds.tryRead(delegate, docs, offset);
if (result != null) {
return result;
}
}
}
return null;
}

BlockLoader.Block tryReadAHead(BlockLoader.BlockFactory factory, BlockLoader.Docs docs, int offset) throws IOException {
if (ords instanceof BaseDenseNumericValues denseOrds && (valuesSorted || entry.termsDictEntry.termsDictSize == 1)) {
int firstDoc = docs.get(offset);
denseOrds.advanceExact(firstDoc);
long startValue = denseOrds.longValue();
Expand Down Expand Up @@ -438,10 +460,18 @@ public TermsEnum termsEnum() throws IOException {
public BlockLoader.Block tryRead(BlockLoader.BlockFactory factory, BlockLoader.Docs docs, int offset) throws IOException {
return null;
}

BlockLoader.Block tryReadAHead(BlockLoader.BlockFactory factory, BlockLoader.Docs docs, int offset) throws IOException {
return null;
}
}

abstract static class BaseDenseNumericValues extends NumericDocValues implements BlockLoader.OptionalColumnAtATimeReader {
abstract long lookAheadValueAt(int targetDoc) throws IOException;

BlockLoader.Block tryRead(BlockLoader.SingletonLongBuilder builder, BlockLoader.Docs docs, int offset) throws IOException {
return null;
}
}

abstract static class BaseSortedSetDocValues extends SortedSetDocValues {
Expand Down Expand Up @@ -1256,41 +1286,49 @@ public long longValue() throws IOException {

@Override
public BlockLoader.Block tryRead(BlockLoader.BlockFactory factory, BlockLoader.Docs docs, int offset) throws IOException {
assert maxOrd == -1 : "unexpected maxOrd[" + maxOrd + "]";
try (BlockLoader.SingletonLongBuilder builder = factory.singletonLongs(docs.count() - offset)) {
return tryRead(builder, docs, offset);
}
}

@Override
BlockLoader.Block tryRead(BlockLoader.SingletonLongBuilder builder, BlockLoader.Docs docs, int offset) throws IOException {
final int docsCount = docs.count();
doc = docs.get(docsCount - 1);
try (BlockLoader.SingletonLongBuilder builder = factory.singletonLongs(docs.count() - offset)) {
for (int i = offset; i < docsCount;) {
int index = docs.get(i);
final int blockIndex = index >>> ES819TSDBDocValuesFormat.NUMERIC_BLOCK_SHIFT;
final int blockInIndex = index & ES819TSDBDocValuesFormat.NUMERIC_BLOCK_MASK;
if (blockIndex != currentBlockIndex) {
assert blockIndex > currentBlockIndex : blockIndex + " < " + currentBlockIndex;
// no need to seek if the loading block is the next block
if (currentBlockIndex + 1 != blockIndex) {
valuesData.seek(indexReader.get(blockIndex));
}
currentBlockIndex = blockIndex;
for (int i = offset; i < docsCount;) {
int index = docs.get(i);
final int blockIndex = index >>> ES819TSDBDocValuesFormat.NUMERIC_BLOCK_SHIFT;
final int blockInIndex = index & ES819TSDBDocValuesFormat.NUMERIC_BLOCK_MASK;
if (blockIndex != currentBlockIndex) {
assert blockIndex > currentBlockIndex : blockIndex + " < " + currentBlockIndex;
// no need to seek if the loading block is the next block
if (currentBlockIndex + 1 != blockIndex) {
valuesData.seek(indexReader.get(blockIndex));
}
currentBlockIndex = blockIndex;
if (bitsPerOrd == -1) {
decoder.decode(valuesData, currentBlock);
} else {
decoder.decodeOrdinals(valuesData, currentBlock, bitsPerOrd);
}
}

// Try to append more than just one value:
// Instead of iterating over docs and find the max length, take an optimistic approach to avoid as
// many comparisons as there are remaining docs and instead do at most 7 comparisons:
int length = 1;
int remainingBlockLength = Math.min(ES819TSDBDocValuesFormat.NUMERIC_BLOCK_SIZE - blockInIndex, docsCount - i);
for (int newLength = remainingBlockLength; newLength > 1; newLength = newLength >> 1) {
int lastIndex = i + newLength - 1;
if (isDense(index, docs.get(lastIndex), newLength)) {
length = newLength;
break;
}
// Try to append more than just one value:
// Instead of iterating over docs and find the max length, take an optimistic approach to avoid as
// many comparisons as there are remaining docs and instead do at most 7 comparisons:
int length = 1;
int remainingBlockLength = Math.min(ES819TSDBDocValuesFormat.NUMERIC_BLOCK_SIZE - blockInIndex, docsCount - i);
for (int newLength = remainingBlockLength; newLength > 1; newLength = newLength >> 1) {
int lastIndex = i + newLength - 1;
if (isDense(index, docs.get(lastIndex), newLength)) {
length = newLength;
break;
}
builder.appendLongs(currentBlock, blockInIndex, length);
i += length;
}
return builder.build();
builder.appendLongs(currentBlock, blockInIndex, length);
i += length;
}
return builder.build();
}

@Override
Expand Down Expand Up @@ -1624,4 +1662,59 @@ private static class TermsDictEntry {
int maxBlockLength;
}

static final class SingletonLongToSingletonOrdinalDelegate implements BlockLoader.SingletonLongBuilder {
private final BlockLoader.SingletonOrdinalsBuilder builder;

SingletonLongToSingletonOrdinalDelegate(BlockLoader.SingletonOrdinalsBuilder builder) {
this.builder = builder;
}

@Override
public BlockLoader.SingletonLongBuilder appendLong(long value) {
throw new UnsupportedOperationException();
}

@Override
public BlockLoader.SingletonLongBuilder appendLongs(long[] values, int from, int length) {
// Unfortunately, no array copy here...
// Since we need to loop here, let's also keep track of min/max.
int minOrd = Integer.MAX_VALUE;
int maxOrd = Integer.MIN_VALUE;
int counter = 0;
int[] convertedOrds = new int[length];
int end = from + length;
for (int j = from; j < end; j++) {
int ord = Math.toIntExact(values[j]);
convertedOrds[counter++] = ord;
minOrd = Math.min(minOrd, ord);
maxOrd = Math.max(maxOrd, ord);
}
builder.appendOrds(convertedOrds, 0, length, minOrd, maxOrd);
return this;
}

@Override
public BlockLoader.Block build() {
return builder.build();
}

@Override
public BlockLoader.Builder appendNull() {
throw new UnsupportedOperationException();
}

@Override
public BlockLoader.Builder beginPositionEntry() {
throw new UnsupportedOperationException();
}

@Override
public BlockLoader.Builder endPositionEntry() {
throw new UnsupportedOperationException();
}

@Override
public void close() {}
}

}
2 changes: 1 addition & 1 deletion server/src/main/java/org/elasticsearch/index/mapper/BlockDocValuesReader.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -756,7 +756,7 @@ public BlockLoader.Block read(BlockFactory factory, Docs docs, int offset) throw
return block;
}
}
try (var builder = factory.singletonOrdinalsBuilder(ordinals, docs.count() - offset)) {
try (var builder = factory.singletonOrdinalsBuilder(ordinals, docs.count() - offset, false)) {
for (int i = offset; i < docs.count(); i++) {
int doc = docs.get(i);
if (doc < ordinals.docID()) {
Expand Down
4 changes: 3 additions & 1 deletion server/src/main/java/org/elasticsearch/index/mapper/BlockLoader.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -447,7 +447,7 @@ interface BlockFactory {
/**
* Build a reader for reading {@link SortedDocValues}
*/
SingletonOrdinalsBuilder singletonOrdinalsBuilder(SortedDocValues ordinals, int count);
SingletonOrdinalsBuilder singletonOrdinalsBuilder(SortedDocValues ordinals, int count, boolean isDense);

/**
* Build a reader for reading {@link SortedSetDocValues}
Expand Down Expand Up @@ -548,6 +548,8 @@ interface SingletonOrdinalsBuilder extends Builder {
* Appends an ordinal to the builder.
*/
SingletonOrdinalsBuilder appendOrd(int value);

SingletonOrdinalsBuilder appendOrds(int[] values, int from, int length, int minOrd, int maxOrd);
}

interface SortedSetOrdinalsBuilder extends Builder {
Expand Down
Loading