Skip to content

Refine indexing pressure accounting in semantic bulk inference filter #129320

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

Open
wants to merge 7 commits into
base: main
Choose a base branch
from
Open

Refine indexing pressure accounting in semantic bulk inference filter #129320

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
e44d8c7
Refine indexing pressure accounting in semantic bulk inference filter
jimczi Jun 12, 2025
4c56277
Update docs/changelog/129320.yaml
jimczi Jun 12, 2025
aeda08b
fix uts
jimczi Jun 12, 2025
a12c9f1
Merge remote-tracking branch 'origin/indexing_pressure_bulk_inference…
jimczi Jun 12, 2025
c85718d
Address review comments and copy the source in place when possible
jimczi Jun 12, 2025
474c9a7
Merge branch 'main' into indexing_pressure_bulk_inference
jimczi Jun 12, 2025
6e62f09
Merge branch 'main' into indexing_pressure_bulk_inference
jimczi Jun 13, 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
5 changes: 5 additions & 0 deletions docs/changelog/129320.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
pr: 129320
summary: Refine indexing pressure accounting in semantic bulk inference filter
area: Relevance
type: enhancement
issues: []
235 changes: 134 additions & 101 deletions .../java/org/elasticsearch/xpack/inference/action/filter/ShardBulkInferenceActionFilter.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,9 @@
import org.elasticsearch.cluster.metadata.InferenceFieldMetadata;
import org.elasticsearch.cluster.metadata.ProjectMetadata;
import org.elasticsearch.cluster.service.ClusterService;
import org.elasticsearch.common.bytes.BytesArray;
import org.elasticsearch.common.bytes.BytesReference;
import org.elasticsearch.common.bytes.CompositeBytesReference;
import org.elasticsearch.common.settings.Setting;
import org.elasticsearch.common.unit.ByteSizeValue;
import org.elasticsearch.common.util.concurrent.AtomicArray;
Expand All @@ -52,6 +54,7 @@
import org.elasticsearch.tasks.Task;
import org.elasticsearch.xcontent.XContent;
import org.elasticsearch.xcontent.XContentBuilder;
import org.elasticsearch.xcontent.XContentFactory;
import org.elasticsearch.xcontent.XContentParser;
import org.elasticsearch.xcontent.XContentParserConfiguration;
import org.elasticsearch.xcontent.XContentType;
Expand Down Expand Up @@ -469,17 +472,17 @@ private void recordRequestCountMetrics(Model model, int incrementBy, Throwable t
* Adds all inference requests associated with their respective inference IDs to the given {@code requestsMap}
* for the specified {@code item}.
*
* @param item The bulk request item to process.
* @param itemIndex The position of the item within the original bulk request.
* @param item The bulk request item to process.
* @param itemIndex The position of the item within the original bulk request.
* @param requestsMap A map storing inference requests, where each key is an inference ID,
* and the value is a list of associated {@link FieldInferenceRequest} objects.
* @return The total content length of all newly added requests, or {@code 0} if no requests were added.
*/
private long addFieldInferenceRequests(BulkItemRequest item, int itemIndex, Map<String, List<FieldInferenceRequest>> requestsMap) {
boolean isUpdateRequest = false;
final IndexRequestWithIndexingPressure indexRequest;
final IndexRequest indexRequest;
if (item.request() instanceof IndexRequest ir) {
indexRequest = new IndexRequestWithIndexingPressure(ir);
indexRequest = ir;
} else if (item.request() instanceof UpdateRequest updateRequest) {
isUpdateRequest = true;
if (updateRequest.script() != null) {
Expand All @@ -493,13 +496,13 @@ private long addFieldInferenceRequests(BulkItemRequest item, int itemIndex, Map<
);
return 0;
}
indexRequest = new IndexRequestWithIndexingPressure(updateRequest.doc());
indexRequest = updateRequest.doc();
} else {
// ignore delete request
return 0;
}

final Map<String, Object> docMap = indexRequest.getIndexRequest().sourceAsMap();
final Map<String, Object> docMap = indexRequest.sourceAsMap();
long inputLength = 0;
for (var entry : fieldInferenceMap.values()) {
String field = entry.getName();
Expand Down Expand Up @@ -535,10 +538,6 @@ private long addFieldInferenceRequests(BulkItemRequest item, int itemIndex, Map<
* This ensures that the field is treated as intentionally cleared,
* preventing any unintended carryover of prior inference results.
*/
if (incrementIndexingPressure(indexRequest, itemIndex) == false) {
return inputLength;
}

var slot = ensureResponseAccumulatorSlot(itemIndex);
slot.addOrUpdateResponse(
new FieldInferenceResponse(field, sourceField, null, order++, 0, null, EMPTY_CHUNKED_INFERENCE)
Expand Down Expand Up @@ -578,10 +577,6 @@ private long addFieldInferenceRequests(BulkItemRequest item, int itemIndex, Map<
List<FieldInferenceRequest> requests = requestsMap.computeIfAbsent(inferenceId, k -> new ArrayList<>());
int offsetAdjustment = 0;
for (String v : values) {
if (incrementIndexingPressure(indexRequest, itemIndex) == false) {
return inputLength;
}

if (v.isBlank()) {
slot.addOrUpdateResponse(
new FieldInferenceResponse(field, sourceField, v, order++, 0, null, EMPTY_CHUNKED_INFERENCE)
Expand All @@ -604,50 +599,6 @@ private long addFieldInferenceRequests(BulkItemRequest item, int itemIndex, Map<
return inputLength;
}

private static class IndexRequestWithIndexingPressure {
private final IndexRequest indexRequest;
private boolean indexingPressureIncremented;

private IndexRequestWithIndexingPressure(IndexRequest indexRequest) {
this.indexRequest = indexRequest;
this.indexingPressureIncremented = false;
}

private IndexRequest getIndexRequest() {
return indexRequest;
}

private boolean isIndexingPressureIncremented() {
return indexingPressureIncremented;
}

private void setIndexingPressureIncremented() {
this.indexingPressureIncremented = true;
}
}

private boolean incrementIndexingPressure(IndexRequestWithIndexingPressure indexRequest, int itemIndex) {
boolean success = true;
if (indexRequest.isIndexingPressureIncremented() == false) {
try {
// Track operation count as one operation per document source update
coordinatingIndexingPressure.increment(1, indexRequest.getIndexRequest().source().ramBytesUsed());
indexRequest.setIndexingPressureIncremented();
} catch (EsRejectedExecutionException e) {
addInferenceResponseFailure(
itemIndex,
new InferenceException(
"Insufficient memory available to update source on document [" + indexRequest.getIndexRequest().id() + "]",
e
)
);
success = false;
}
}

return success;
}

private FieldInferenceResponseAccumulator ensureResponseAccumulatorSlot(int id) {
FieldInferenceResponseAccumulator acc = inferenceResults.get(id);
if (acc == null) {
Expand Down Expand Up @@ -723,70 +674,152 @@ private void applyInferenceResponses(BulkItemRequest item, FieldInferenceRespons
);
inferenceFieldsMap.put(fieldName, result);
}

BytesReference originalSource = indexRequest.source();
if (useLegacyFormat) {
var newDocMap = indexRequest.sourceAsMap();
for (var entry : inferenceFieldsMap.entrySet()) {
SemanticTextUtils.insertValue(entry.getKey(), newDocMap, entry.getValue());
}
indexRequest.source(newDocMap, indexRequest.getContentType());
XContentBuilder builder = XContentFactory.contentBuilder(indexRequest.getContentType());
builder.map(newDocMap);
var newSource = BytesReference.bytes(builder);
if (incrementIndexingPressure(item, indexRequest, newSource.length())) {
indexRequest.source(newSource, indexRequest.getContentType());
}
} else {
updateSourceWithInferenceFields(item, indexRequest, inferenceFieldsMap);
}
}

/**
* Updates the {@link IndexRequest}'s source to include additional inference fields.
* <p>
* If the original source uses an array-backed {@link BytesReference}, this method attempts an in-place update,
* reusing the existing array where possible and appending additional bytes only if needed.
* <p>
* If the original source is not array-backed, the entire source is replaced with the new source that includes
* the inference fields. In this case, the full size of the new source is accounted for in indexing pressure.
* <p>
* Note: We do not subtract the indexing pressure of the original source since its bytes may be pooled and not
* reclaimable by the garbage collector during the request lifecycle.
*
* @param item The {@link BulkItemRequest} being processed.
* @param indexRequest The {@link IndexRequest} whose source will be updated.
* @param inferenceFieldsMap A map of additional fields to append to the source.
* @throws IOException if building the new source fails.
*/
private void updateSourceWithInferenceFields(
BulkItemRequest item,
IndexRequest indexRequest,
Map<String, Object> inferenceFieldsMap
) throws IOException {
var originalSource = indexRequest.source();
final BytesReference newSource;

// Build a new source by appending the inference fields to the existing source.
try (XContentBuilder builder = XContentBuilder.builder(indexRequest.getContentType().xContent())) {
appendSourceAndInferenceMetadata(builder, originalSource, indexRequest.getContentType(), inferenceFieldsMap);
newSource = BytesReference.bytes(builder);
}

// Calculate the additional size to account for in indexing pressure.
final int additionalSize = originalSource.hasArray() ? newSource.length() - originalSource.length() : newSource.length();

// If we exceed the indexing pressure limit, do not proceed with the update.
if (incrementIndexingPressure(item, indexRequest, additionalSize) == false) {
return;
}

// Apply the updated source to the index request.
if (originalSource.hasArray()) {
// If the original source is backed by an array, perform in-place update:
// - Copy as much of the new source as fits into the original array.
System.arraycopy(
newSource.array(),
newSource.arrayOffset(),
originalSource.array(),
originalSource.arrayOffset(),
originalSource.length()
);

int remainingSize = newSource.length() - originalSource.length();
if (remainingSize > 0) {
// If there are additional bytes, append them as a new BytesArray segment.
byte[] remainingBytes = new byte[remainingSize];
System.arraycopy(
newSource.array(),
newSource.arrayOffset() + originalSource.length(),
remainingBytes,
0,
remainingSize
);
indexRequest.source(
CompositeBytesReference.of(originalSource, new BytesArray(remainingBytes)),
indexRequest.getContentType()
);
} else {
// No additional bytes; just adjust the slice length.
indexRequest.source(originalSource.slice(0, newSource.length()));
}
} else {
try (XContentBuilder builder = XContentBuilder.builder(indexRequest.getContentType().xContent())) {
appendSourceAndInferenceMetadata(builder, indexRequest.source(), indexRequest.getContentType(), inferenceFieldsMap);
indexRequest.source(builder);
// If the original source is not array-backed, replace it entirely.
indexRequest.source(newSource, indexRequest.getContentType());
}
}

/**
* Appends the original source and the new inference metadata field directly to the provided
* {@link XContentBuilder}, avoiding the need to materialize the original source as a {@link Map}.
*/
private void appendSourceAndInferenceMetadata(
XContentBuilder builder,
BytesReference source,
XContentType xContentType,
Map<String, Object> inferenceFieldsMap
) throws IOException {
builder.startObject();

// append the original source
try (
XContentParser parser = XContentHelper.createParserNotCompressed(XContentParserConfiguration.EMPTY, source, xContentType)
) {
// skip start object
parser.nextToken();
while (parser.nextToken() != XContentParser.Token.END_OBJECT) {
builder.copyCurrentStructure(parser);
}
}
long modifiedSourceSize = indexRequest.source().ramBytesUsed();

// Add the indexing pressure from the source modifications.
// Don't increment operation count because we count one source update as one operation, and we already accounted for those
// in addFieldInferenceRequests.
// add the inference metadata field
builder.field(InferenceMetadataFieldsMapper.NAME);
try (XContentParser parser = XContentHelper.mapToXContentParser(XContentParserConfiguration.EMPTY, inferenceFieldsMap)) {
builder.copyCurrentStructure(parser);
}

builder.endObject();
}

private boolean incrementIndexingPressure(BulkItemRequest item, IndexRequest indexRequest, int inc) {
try {
coordinatingIndexingPressure.increment(0, modifiedSourceSize - originalSource.ramBytesUsed());
if (inc > 0) {
coordinatingIndexingPressure.increment(1, inc);
}
return true;
} catch (EsRejectedExecutionException e) {
indexRequest.source(originalSource, indexRequest.getContentType());
inferenceStats.bulkRejection().incrementBy(1);
item.abort(
item.index(),
new InferenceException(
"Insufficient memory available to insert inference results into document [" + indexRequest.id() + "]",
"Unable to insert inference results into document ["
+ indexRequest.id()
+ "] due to memory pressure. Please retry the bulk request with fewer documents or smaller document sizes.",
e
)
);
return false;
}
}
}

/**
* Appends the original source and the new inference metadata field directly to the provided
* {@link XContentBuilder}, avoiding the need to materialize the original source as a {@link Map}.
*/
private static void appendSourceAndInferenceMetadata(
XContentBuilder builder,
BytesReference source,
XContentType xContentType,
Map<String, Object> inferenceFieldsMap
) throws IOException {
builder.startObject();

// append the original source
try (XContentParser parser = XContentHelper.createParserNotCompressed(XContentParserConfiguration.EMPTY, source, xContentType)) {
// skip start object
parser.nextToken();
while (parser.nextToken() != XContentParser.Token.END_OBJECT) {
builder.copyCurrentStructure(parser);
}
}

// add the inference metadata field
builder.field(InferenceMetadataFieldsMapper.NAME);
try (XContentParser parser = XContentHelper.mapToXContentParser(XContentParserConfiguration.EMPTY, inferenceFieldsMap)) {
builder.copyCurrentStructure(parser);
}

builder.endObject();
}

static IndexRequest getIndexRequestOrNull(DocWriteRequest<?> docWriteRequest) {
if (docWriteRequest instanceof IndexRequest indexRequest) {
return indexRequest;
Expand Down
8 changes: 7 additions & 1 deletion ...n/inference/src/main/java/org/elasticsearch/xpack/inference/telemetry/InferenceStats.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,12 @@
import java.util.Map;
import java.util.Objects;

public record InferenceStats(LongCounter requestCount, LongHistogram inferenceDuration) {
public record InferenceStats(LongCounter requestCount, LongHistogram inferenceDuration, LongCounter bulkRejection) {

public InferenceStats {
Objects.requireNonNull(requestCount);
Objects.requireNonNull(inferenceDuration);
Objects.requireNonNull(bulkRejection);
}

public static InferenceStats create(MeterRegistry meterRegistry) {
Expand All @@ -38,6 +39,11 @@ public static InferenceStats create(MeterRegistry meterRegistry) {
"es.inference.requests.time",
"Inference API request counts for a particular service, task type, model ID",
"ms"
),
meterRegistry.registerLongCounter(
"es.inference.bulk.rejection.total",
"Count of bulk request rejections for semantic text processing due to insufficient available memory",
"operations"
)
);
}
Expand Down
2 changes: 1 addition & 1 deletion ...t/java/org/elasticsearch/xpack/inference/action/BaseTransportInferenceActionTestCase.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,7 +88,7 @@ public void setUp() throws Exception {
licenseState = mock();
modelRegistry = mock();
serviceRegistry = mock();
inferenceStats = new InferenceStats(mock(), mock());
inferenceStats = new InferenceStats(mock(), mock(), mock());
streamingTaskManager = mock();

action = createAction(
Expand Down
Loading
Loading