What's Changed

• Regenerate client by @flobernd in #8990
  • DenseVectorIndexOptions gains the bbq_disk properties AutoCalibrate, Bits, ClusterSize,
    DefaultVisitPercentage and Precondition.
  • DenseVectorIndexOptions.ConfidenceInterval is now marked [Obsolete] (deprecated since 9.5.0).
  • SourceFieldMode gains ColumnarStored (columnar_stored), so index templates and mappings that use the new mode
    no longer fail deserialization.
  • Breaking: InnerRetriever.Normalizer and InnerRetriever.Weight are now optional (ScoreNormalizer? and
    float? instead of required), following the spec change that made normalizer and weight optional on the linear
    retriever's inner retrievers. The [SetsRequiredMembers] constructor no longer takes normalizer and weight
    arguments.
  • Breaking: RoleDescriptor.Global, QueryRole.Global and RoleDescriptorRead.Global now use a single
    GlobalPrivilege instead of a collection, matching the server-side parser, which only accepts an object for the
    global privilege. The Global descriptor methods replace their params array overloads with a single-object
    overload.

Full Changelog: 9.5.1...9.5.2

What's Changed

• Regenerate client by @flobernd in #8957 and #8981
  • Transform destinations now accept aliases, each with an optional move_on_creation flag, through the new DestinationAlias type.
  • Request converter sources regenerated to emit the full client invocation.
  • Breaking: KnnRetriever.NumCandidates is now optional (int? instead of required int), following the spec change that made num_candidates optional on the kNN retriever. The [SetsRequiredMembers] constructor no longer takes a numCandidates argument.
  • Breaking: TransformSummary.Dest and UpdateTransformResponse.Dest now use Elastic.Clients.Elasticsearch.TransformManagement.Destination instead of Elastic.Clients.Elasticsearch.Core.Reindex.Destination.
• Restore delegate caching in IdResolver by @fresh55 in #8973
  • The getter delegate was only cached when resolution failed, so it was rebuilt on every successful call.
  • 100k resolves of the same type improve from 64 ms / 55.2 MB to 3.5 ms / 40 bytes. Partially addresses (#7664).

Full Changelog: 9.5.0...9.5.1

What's Changed

Initial 9.5.x release.

• Regenerate client against the Elasticsearch 9.5 specification by @flobernd in #8930 and #8944
  • New experimental client.Encryption.Reset endpoint, plus a ProjectEncryptionKey indicator on the health report
  • New experimental ES|QL data federation APIs: PutDataSource / GetDataSource / DeleteDataSource and PutDataset / GetDataset / DeleteDataset, with the supporting ESQLDataSource, ESQLDataset and DatasetMapping types
  • New experimental Inference region policy APIs: PutRegionPolicy / GetRegionPolicy / DeleteRegionPolicy
  • New _slice routing parameter (RouteSlice) across the search, get, index, delete, update and bulk APIs, as well as on the multi-search header and the reindex destination
  • FrequentItemSets aggregations now round-trip correctly. The generated variant used an empty discriminator where the converter expected frequent_item_sets, so the property always read back as null
  • Role-template queries in index privileges now deserialize to RoleTemplateQuery. Previously any object-shaped query was unconditionally read as a regular QueryDsl.Query
  • SpanGap and RoleMappingRule.Field variant accessors now return null instead of a default KeyValuePair when the union holds a different variant
  • New enum members: EmbeddingContentType.Audio / .Pdf / .Video, AnthropicTaskType.ChatCompletion, OpenAITaskType.Embedding, and the ClusterPrivilege.ManageReindex / .MonitorReindex values
  • Breaking (mapping): DocValues on 32 property types changes from bool? to Union<bool, DocValuesConfig>? to carry the new doc_values.multi_value setting. The parameterless DocValues() descriptor overload is gone, and reading the property into a bool? no longer compiles. SourceField.Compress and CompressThreshold are removed, and TextIndexPrefixes.MinChars / MaxChars become optional (int?), dropping the two-argument constructor
  • Breaking (ES|QL): the named-parameter arm of Params on EsqlQuery and AsyncQuery changes from KeyValuePair<string, ICollection<FieldValue>> to KeyValuePair<string, Union<ICollection<FieldValue>, ClassifiedNamedParameter>>, which affects every existing named-parameter call site. Both requests also gain a Settings body property
  • Breaking (inference): Rerank now accepts the expanded multimodal request format, so Input becomes Union<ICollection<string>, ICollection<RerankInputObject>>, Query becomes Union<string, RerankInputObject>, and the Input(params string[]) overload is removed. EmbeddingContentObject.Content becomes ICollection<EmbeddingContentObjectItem> and EmbeddingContentObjectContents is gone, OpenAIServiceSettings.ApiKey is now optional, and JinaAIServiceSettings.ElementType is renamed to EmbeddingType
  • Breaking (snapshot): SourceOnlyRepositorySettings is replaced by the ISourceOnlyRepositorySettings variant family with per-backend settings types, constructed through SourceOnlyRepositorySettingsFactory (ForAzure, ForGcs, ForReadOnlyUrl, ForS3, ForSharedFileSystem). SourceOnlyRepository.Settings now takes the interface, and its descriptor overloads are replaced by a factory callback
  • Breaking (security): RoleInfo.Indices changes from IndicesPrivileges to IndicesPrivilegesRead, and role definitions gain DataSourcePrivileges
  • Breaking (stats): GetStats.Time is renamed to GetTime, ShardStats.Indices becomes a dictionary of ShardStats, the DenseVectorOffHeapStats.TotalCenifSize typo is corrected to TotalCenivfSize, and newly required members change the positional constructors of IndexingStats, SearchStats, FlushStats, MappingStats and MachineLearning.Defaults
• Make custom JSON converters safe for multi-segment readers by @flobernd in #8923
  • The hand-written date, duration, stringified and fractional numeric converters no longer fail when the Utf8JsonReader is constructed over a multi-segment ReadOnlySequence<byte>, which unblocks sequence-based deserializers such as a streaming reader over a PipeReader
  • The existing single-segment typed deserialization path is unchanged

New Contributors

• @charlotte-hoblik made their first contribution in #8901

Full Changelog: 9.4.2...9.5.0

What's Changed

• Add plugin-defined (non-exhaustive) variant registry by @flobernd in #8911
  • Unknown discriminators on variant union families (properties, analyzers, char filters, token filters, tokenizers, repositories, similarity settings, aggregations, queries, etc.) now deserialize into an Unknown* carrier and re-serialize losslessly — no data is dropped for plugin-defined or future variants the client doesn't model
  • Opt into strong typing by registering a CLR type for a discriminator: settings.Variants.Register<IProperty, MyPluginProperty>("my_plugin"); registrations are scoped per client instance and AOT-safe — see the Plugin-defined variant types documentation for details
• Regenerate client by @flobernd in #8908 and #8917
  • Generated Unknown* carrier types for all variant families (properties, analyzers, char filters, token filters, tokenizers, repositories, similarity settings, aggregations, queries, suggesters, rescorers) to support the non-exhaustive variant registry
  • Improved XML documentation coverage across generated client methods, request/response types, and type members that previously had no IntelliSense descriptions

Full Changelog: 9.4.1...9.4.2

What's Changed

• Regenerate client by @flobernd in #8899 and #8907
  • Fixes (de-)serialization of unions with three or more variants — previously, hand-rolled union converters could fail to select the correct variant
  • Search response shape fix in InnerHits: the Fields property is split into Field (single-field selector using Fields) and a new Fields collection of FieldAndFormat — existing usages of InnerHits.Fields = … may need to be retargeted to Field
  • DataStreamLifecycle gains EffectiveRetention and RetentionDeterminedBy; new GlobalRetention and RetentionSource types; GetDataLifecycleResponse now exposes GlobalRetention
  • Reindex rethrottle response now models parent task progress via the new ParentReindexStatus type
  • Inference and _mvt content-type alignment: chat_completion_unified / stream_completion send Content-Type: application/json; _mvt uses the versioned application/vnd.elasticsearch+vnd.mapbox-vector-tile Accept header

Full Changelog: 9.4.0...9.4.1

What's Changed

Initial 9.4.x release.

Full Changelog: 9.3.7...9.4.0

What's Changed

• Regenerate client by @flobernd in #8898
  • Set explicit ContentType: application/json for ChatCompletionUnified and StreamCompletion streaming inference requests
  • Switch SearchMvt Accept header to application/vnd.elasticsearch+vnd.mapbox-vector-tile (Elasticsearch product-flavored content-type)
  • Bump Elastic.Transport to 0.17.1
  • Document default max_page_search_size values for pivot (500) and latest (5000) transforms

Full Changelog: 9.3.6...9.3.7

What's Changed

• Remove obsolete ElasticsearchClientProductRegistration overrides by @flobernd in #8891
  • Fixes #8890

Full Changelog: 9.3.5...9.3.6

Overview

• 1. Binary-response endpoints now return ElasticsearchStreamResponse
• 1. Default Accept header on SearchMvtRequest
• 1. ES|QL QueryAsStreamAsync return type

Breaking changes

1. Binary-response endpoints now return ElasticsearchStreamResponse

Impact : Medium.

Endpoints whose response bodies are binary ( search_mvt ) or that the client intentionally exposes as streams (ES|QL, streaming Inference) now return Elastic.Transport.Products.Elasticsearch.ElasticsearchStreamResponse instead of endpoint-specific response classes.

The following response types have been removed:

• Elastic.Clients.Elasticsearch.SearchMvtResponse

• Elastic.Clients.Elasticsearch.Esql.EsqlQueryResponse

• Elastic.Clients.Elasticsearch.Esql.AsyncQueryResponse

• Elastic.Clients.Elasticsearch.Esql.AsyncQueryGetResponse

• Elastic.Clients.Elasticsearch.Esql.AsyncQueryStopResponse

• Elastic.Clients.Elasticsearch.Inference.ChatCompletionUnifiedResponse

• Elastic.Clients.Elasticsearch.Inference.StreamCompletionResponse

All sync and async overloads of the following client methods now return ElasticsearchStreamResponse / Task<ElasticsearchStreamResponse> :

• ElasticsearchClient.SearchMvt / SearchMvtAsync

• EsqlNamespacedClient.Query / QueryAsync

• EsqlNamespacedClient.AsyncQuery / AsyncQueryAsync

• EsqlNamespacedClient.AsyncQueryGet / AsyncQueryGetAsync

• EsqlNamespacedClient.AsyncQueryStop / AsyncQueryStopAsync

• InferenceNamespacedClient.ChatCompletionUnified / ChatCompletionUnifiedAsync

• InferenceNamespacedClient.StreamCompletion / StreamCompletionAsync

Replace response.Data (previously a byte[] ) with response.Body ( System.IO.Stream ). The response owns the stream, so disposing the response disposes the stream — no separate disposal of Body is required.

// 9.3.4 and below SearchMvtResponse response = await client.SearchMvtAsync(req); byte[] tile = response.Data;

// 9.3.5 using var response = await client.SearchMvtAsync(req); Stream tile = response.Body;

Elastic.Clients.Elasticsearch now requires Elastic.Transport >= 0.16.0 (previously 0.15.1 ); ElasticsearchStreamResponse was introduced in transport 0.16.0.

2. Default Accept header on SearchMvtRequest

Impact : Low.

SearchMvtRequest now overrides the inherited IRequestConfiguration? RequestConfiguration property. The getter returns a static default ( new RequestConfiguration { Accept = "application/vnd.mapbox-vector-tile" } ) when the user has not assigned a value. The setter still forwards to the base class, so user-assigned values take precedence.

Code that read request.RequestConfiguration and expected null when nothing was assigned will now observe the default singleton instead.

No Accept override is emitted for ES|QL or Inference request types; the transport default ( application/json ) continues to apply to them.

3. ES|QL QueryAsStreamAsync return type

Impact : Low.

EsqlNamespacedClient.QueryAsStreamAsync (both the generic <TDocument> and the non-generic overload) return type changed from Task<Elastic.Transport.StreamResponse> to Task<Elastic.Transport.Products.Elasticsearch.ElasticsearchStreamResponse> . ElasticsearchStreamResponse is a drop-in replacement with the same Body ( System.IO.Stream ) property, additionally exposing IsValidResponse , ElasticsearchServerError , ElasticsearchWarnings , ApiCallDetails , and DebugInformation .

Overview

• 1. Improved usability of CompositeAggregation

Breaking changes

1. Improved usability of CompositeAggregation

Impact : Low.

The type of the Sources property has been changed from ICollection<IDictionary<string, CompositeAggregationSource>> to ICollection<KeyValuePair<string, CompositeAggregationSource>> . This corresponds to the Elasticsearch standard for modeling ordered dictionaries in the REST API.

CompositeAggregationSource is now also a container (similar to Aggregation , Query , etc.). This change improves usability due to specialized code generation. For example, implicit conversion operators from all existing variants ( CompositeTermsAggregation , CompositeHistogramAggregation , etc.) to CompositeAggregationSource are now generated.

As a consequence, the object initializer syntax changes as follows:

// 9.1.7 and below

var request = new SearchRequest { Aggregations = new Dictionary<string, Aggregation> { { "my_composite", new CompositeAggregation { Sources = [ new Dictionary<string, CompositeAggregationSource> { { "my_terms", new CompositeAggregationSource { Terms = new CompositeTermsAggregation { // ... } }} }, new Dictionary<string, CompositeAggregationSource> { { "my_histo", new CompositeAggregationSource { Histogram = new CompositeHistogramAggregation(0.5) { // ... } }} } ] }} } };

// 9.1.8 and above

var request = new SearchRequest { Aggregations = new Dictionary<string, Aggregation> { { "my_composite", new CompositeAggregation { Sources = [ new KeyValuePair<string, CompositeAggregationSource>( "my_terms", new CompositeTermsAggregation { // ... } ), new KeyValuePair<string, CompositeAggregationSource>( "my_histo", new CompositeHistogramAggregation(0.5) { // ... } ) ] }} } };

• Implicit conversion from CompositeTermsAggregation to CompositeAggregationSource .

• Implicit conversion from CompositeHistogramAggregation to CompositeAggregationSource .

In addition, this change allows optimized Fluent syntax to be generated, which ultimately avoids a previously existing ambiguity:

// 9.1.7 and below

var descriptor = new SearchRequestDescriptor() .Aggregations(aggs => aggs .Add("my_composize", agg => agg .Composite(composite => composite .Sources( x => x.Add("my_terms", x => x.Terms(/* ... /)), x => x.Add("my_histo", x => x.Histogram(/ ... */)) ) ) ) );

// 9.1.8 and above

var descriptor = new SearchRequestDescriptor() .Aggregations(aggs => aggs .Add("my_composize", agg => agg .Composite(composite => composite .Sources(sources => sources .Add("my_terms", x => x.Terms(/* ... /)) .Add("my_histo", x => x.Histogram(/ ... */)) ) ) ) );

• This is the params overload that initializes the Sources collection with multiple entries.

• Add dictionary item 1 that contains a single Terms aggregation entry.

• Add dictionary item 2 that contains a single Histogram aggregation entry.

The old syntax was tricky because the 9.1.8 example also compiled successfully, but the .Add referred to the first dictionary both times. This ultimately resulted in a list with only one dictionary, which had multiple entries, and thus an invalid request.

Overview

• 1. Improved usability of Percentiles aggregation results

Breaking changes

1. Improved usability of Percentiles aggregation results

Impact : Low.

The type of the Values property for the following classes got changed from the Percentiles union to a simple PercentilesItem collection:

• HdrPercentilesAggregate

• HdrPercentileRanksAggregate

• TDigestPercentilesAggregate

• TDigestPercentileRanksAggregate

• PercentilesBucketAggregate

The Percentiles union type got removed as it's no longer used.

Overview

• 1. Container types
• 1. Removal of certain generic request descriptors
• 1. Removal of certain descriptor constructors and related request APIs
• 1. Date / Time / Duration values
• 1. ExtendedBounds
• 1. Field.Format
• 1. Field / Fields semantics
• 1. FieldValue
• 1. FieldSort
• 1. Descriptor types class -> struct

Breaking changes

1. Container types

Impact : High.

Container types now use regular properties for their variants instead of static factory methods ( read more ).

This change primarily affects the Query and Aggregation types.

// 8.x new SearchRequest { Query = Query.MatchAll( new MatchAllQuery { } ) };

// 9.0 new SearchRequest { Query = new Query { MatchAll = new MatchAllQuery { } } };

Previously required methods like e.g. TryGet<TVariant>(out) have been removed.

The new recommended way of inspecting container types is to use simple pattern matching:

var query = new Query();

if (query.Nested is { } nested) { // We have a nested query. }

2. Removal of certain generic request descriptors

Impact : High.

Removed the generic version of some request descriptors for which the corresponding requests do not contain inferrable properties.

These descriptors were generated unintentionally.

When migrating, the generic type parameter must be removed from the type, e.g., AsyncSearchStatusRequestDescriptor<TDocument> should become just AsyncSearchStatusRequestDescriptor .

List of affected descriptors:

• AsyncQueryDeleteRequestDescriptor<TDocument>

• AsyncQueryGetRequestDescriptor<TDocument>

• AsyncSearchStatusRequestDescriptor<TDocument>

• DatabaseConfigurationDescriptor<TDocument>

• DatabaseConfigurationFullDescriptor<TDocument>

• DeleteAsyncRequestDescriptor<TDocument>

• DeleteAsyncSearchRequestDescriptor<TDocument>

• DeleteDataFrameAnalyticsRequestDescriptor<TDocument>

• DeleteGeoipDatabaseRequestDescriptor<TDocument>

• DeleteIpLocationDatabaseRequestDescriptor<TDocument>

• DeleteJobRequestDescriptor<TDocument>

• DeletePipelineRequestDescriptor<TDocument>

• DeleteScriptRequestDescriptor<TDocument>

• DeleteSynonymRequestDescriptor<TDocument>

• EqlDeleteRequestDescriptor<TDocument>

• EqlGetRequestDescriptor<TDocument>

• GetAsyncRequestDescriptor<TDocument>

• GetAsyncSearchRequestDescriptor<TDocument>

• GetAsyncStatusRequestDescriptor<TDocument>

• GetDataFrameAnalyticsRequestDescriptor<TDocument>

• GetDataFrameAnalyticsStatsRequestDescriptor<TDocument>

• GetEqlStatusRequestDescriptor<TDocument>

• GetGeoipDatabaseRequestDescriptor<TDocument>

• GetIpLocationDatabaseRequestDescriptor<TDocument>

• GetJobsRequestDescriptor<TDocument>

• GetPipelineRequestDescriptor<TDocument>

• GetRollupCapsRequestDescriptor<TDocument>

• GetRollupIndexCapsRequestDescriptor<TDocument>

• GetScriptRequestDescriptor<TDocument>

• GetSynonymRequestDescriptor<TDocument>

• IndexModifyDataStreamActionDescriptor<TDocument>

• PreprocessorDescriptor<TDocument>

• PutGeoipDatabaseRequestDescriptor<TDocument>

• PutIpLocationDatabaseRequestDescriptor<TDocument>

• PutScriptRequestDescriptor<TDocument>

• PutSynonymRequestDescriptor<TDocument>

• QueryVectorBuilderDescriptor<TDocument>

• RankDescriptor<TDocument>

• RenderSearchTemplateRequestDescriptor<TDocument>

• SmoothingModelDescriptor<TDocument>

• StartDataFrameAnalyticsRequestDescriptor<TDocument>

• StartJobRequestDescriptor<TDocument>

• StopDataFrameAnalyticsRequestDescriptor<TDocument>

• StopJobRequestDescriptor<TDocument>

• TokenizationConfigDescriptor<TDocument>

• UpdateDataFrameAnalyticsRequestDescriptor<TDocument>

3. Removal of certain descriptor constructors and related request APIs

Impact : High.

Removed (TDocument, IndexName) descriptor constructors and related request APIs for all requests with IndexName and Id path parameters.

For example:

// 8.x public IndexRequestDescriptor(TDocument document, IndexName index, Id? id) { } public IndexRequestDescriptor(TDocument document, IndexName index) { } public IndexRequestDescriptor(TDocument document, Id? id) { } public IndexRequestDescriptor(TDocument document) { }

// 9.0 public IndexRequestDescriptor(TDocument document, IndexName index, Id? id) { } public IndexRequestDescriptor(TDocument document, Id? id) { } public IndexRequestDescriptor(TDocument document) { }

These overloads caused invocation ambiguities since both, IndexName and Id implement implicit conversion operators from string .

Alternative with same semantics:

// Descriptor constructor. new IndexRequestDescriptor(document, "my_index", Id.From(document));

// Request API method. await client.IndexAsync(document, "my_index", Id.From(document), ...);

4. Date / Time / Duration values

Impact : High.

In places where previously long or double was used to represent a date/time/duration value, DateTimeOffset or TimeSpan is now used instead.

5. ExtendedBounds

Impact : High.

Removed ExtendedBoundsDate / ExtendedBoundsDateDescriptor , ExtendedBoundsFloat / ExtendedBoundsFloatDescriptor .

Replaced by ExtendedBounds<T> , ExtendedBoundsOfFieldDateMathDescriptor , and ExtendedBoundsOfDoubleDescriptor .

6. Field.Format

Impact : Low.

Removed Field.Format property and corresponding constructor and inferrer overloads.

This property has not been used for some time (replaced by the FieldAndFormat type).

7. Field / Fields semantics

Impact : Low.

Field / Fields static factory methods and conversion operators no longer return nullable references but throw exceptions instead ( Field ) if the input string / Expression / PropertyInfo argument is null .

This makes implicit conversions to Field more user-friendly without requiring the null-forgiveness operator ( ! ) ( read more ).

8. FieldValue

Impact : Low.

Removed FieldValue.IsLazyDocument , FieldValue.IsComposite , and the corresponding members in the FieldValue.ValueKind enum.

These values have not been used for some time.

9. FieldSort

Impact : High.

Removed FieldSort parameterless constructor.

Please use the new constructor instead:

public FieldSort(Elastic.Clients.Elasticsearch.Field field)

Impact : Low.

Removed static FieldSort.Empty member.

Sorting got reworked which makes this member obsolete ( read more ).

10. Descriptor types class -> struct

Impact : Low.

All descriptor types are now implemented as struct instead of class .

What's Changed

• Update Elastic.Transport to 1.1.0 by @flobernd in #8992
  • Records the Elastic-Cluster-Name response header (sent by Elasticsearch 9.6+ when http.headers.cluster_name.enabled is set) as the db.elasticsearch.cluster.name span attribute
  • Fixes a duplicate-key exception when ParseAllHeaders or an overlapping ResponseHeadersToParse was used while OpenTelemetry listeners were active
• Fix routing inference for bulk update and delete operations by @fresh55 in #8995 Fixes #8978
  • Bulk update operations (including the fluent IdFrom path and script upserts) now infer routing from the document instead of omitting it
  • Bulk delete operations no longer emit "routing":null for types without a routing property, and Delete<T>(doc) / DeleteMany<T>(...) keep routing and _index inference
• Avoid closure allocations in RoutingResolver by @fresh55 in #8979 Fixes #7664
  • RoutingResolver.Resolve() no longer allocates on any path (join-field routing drops from 160 B/op to 0 B/op)
• Regenerate client by @flobernd in #9002 and #9006
  • Adds the Certificates (_ssl/certificates), Cluster.RemoteInfo, DanglingIndices.DeleteDanglingIndex / ImportDanglingIndex, Ilm.ExplainLifecycle, Indices.AddBlock and Ml.PreviewDatafeed endpoints
  • Adds the moving_avg pipeline aggregation (Aggregation.MovingAvg with EWMA, Holt, Holt-Winters, linear and simple models)
  • Adds MaxConsecutiveExtractionFailures to DatafeedConfig and the put/update datafeed requests
  • Adds AnalysisStats to DataframeAnalytics

Full Changelog: 8.19.26...8.19.27

What's Changed

• Regenerate client by @flobernd in #8989
  • Breaking: InnerRetriever.Normalizer and InnerRetriever.Weight are now optional (ScoreNormalizer? and
    float? instead of required), following the spec change that made normalizer and weight optional on the linear
    retriever's inner retrievers. The [SetsRequiredMembers] constructor no longer takes normalizer and weight
    arguments.
  • Breaking: RoleDescriptor.Global, QueryRole.Global and RoleDescriptorRead.Global now use a single
    GlobalPrivilege instead of a collection, matching the server-side parser, which only accepts an object for the
    global privilege. The Global descriptor methods replace their params array overloads with a single-object
    overload.

Full Changelog: 8.19.25...8.19.26

What's Changed

• Regenerate client by @flobernd in #8955 and #8982
  • Transform destinations now accept aliases, each with an optional move_on_creation flag, through the new DestinationAlias type.
  • Request converter sources regenerated to emit the full client invocation.
  • Breaking: KnnRetriever.NumCandidates is now optional (int? instead of required int), following the spec change that made num_candidates optional on the kNN retriever. The [SetsRequiredMembers] constructor no longer takes a numCandidates argument.
  • Breaking: TransformSummary.Dest and UpdateTransformResponse.Dest now use Elastic.Clients.Elasticsearch.TransformManagement.Destination instead of Elastic.Clients.Elasticsearch.Core.Reindex.Destination.
• Restore delegate caching in IdResolver by @fresh55 in #8973
  • The getter delegate was only cached when resolution failed, so it was rebuilt on every successful call.
  • 100k resolves of the same type improve from 64 ms / 55.2 MB to 3.5 ms / 40 bytes. Partially addresses (#7664).

Full Changelog: 8.19.24...8.19.25

What's Changed

• Make custom JSON converters safe for multi-segment readers by @flobernd in #8923
  • The hand-written date, duration, stringified and fractional numeric converters no longer fail when the Utf8JsonReader is constructed over a multi-segment ReadOnlySequence<byte>, which unblocks sequence-based deserializers such as a streaming reader over a PipeReader
  • The existing single-segment typed deserialization path is unchanged
• Regenerate client by @flobernd in #8922 and #8945
  • FrequentItemSets aggregations now round-trip correctly. The generated variant used an empty discriminator where the converter expected frequent_item_sets, so the property always read back as null
  • Role-template queries in index privileges now deserialize to RoleTemplateQuery. Previously the union had no discriminator between the two object arms, so any object-shaped query was unconditionally read as a regular QueryDsl.Query
  • SpanQuery.SpanGap now returns null instead of a default KeyValuePair when the union holds a different variant, so SpanGap is not null is no longer always true
  • New basic query parameter on GetTransformStats, which returns a cheaper subset of statistics by skipping operations_behind, changes_last_detected_at, last_search_time and the checkpoint timestamps
  • New settings and response fields: IndexSettings.Unassigned exposing index.unassigned.node_left.delayed_timeout, OnScriptError on runtime field definitions, and Allocations shard-allocation statistics on node stats
  • Breaking: TextIndexPrefixes.MinChars and MaxChars become optional (int?) and the two-argument TextIndexPrefixes(int, int) constructor is removed. This also fixes reading responses that omit either value, which previously failed the required-member check
  • Breaking: JinaAIServiceSettings.ElementType is renamed to EmbeddingType, matching the embedding_type wire name
  • Breaking: NodeInfoDiscover.SeedProviders changes from IReadOnlyCollection<string> to ICollection<string>, and a scalar seed_providers value from the node info API now deserializes instead of throwing

Full Changelog: 8.19.23...8.19.24

What's Changed

• Add plugin-defined (non-exhaustive) variant registry by @flobernd in #8911
  • Unknown discriminators on variant union families (properties, analyzers, char filters, token filters, tokenizers, repositories, similarity settings, aggregations, queries, etc.) now deserialize into an Unknown* carrier and re-serialize losslessly — no data is dropped for plugin-defined or future variants the client doesn't model
  • Opt into strong typing by registering a CLR type for a discriminator: settings.Variants.Register<IProperty, MyPluginProperty>("my_plugin"); registrations are scoped per client instance and AOT-safe — see the Plugin-defined variant types documentation for details
• Regenerate client by @flobernd in #8910 and #8916
  • Generated Unknown* carrier types for all variant families (properties, analyzers, char filters, token filters, tokenizers, repositories, similarity settings, aggregations, queries, suggesters, rescorers) to support the non-exhaustive variant registry
  • Improved XML documentation coverage across generated client methods, request/response types, and type members that previously had no IntelliSense descriptions

Full Changelog: 8.19.22...8.19.23

What's Changed

• Regenerate client by @flobernd in #8897 and #8906
  • Fixes (de-)serialization of unions with three or more variants — previously, hand-rolled union converters could fail to select the correct variant
  • DataStreamLifecycle gains EffectiveRetention and RetentionDeterminedBy; new GlobalRetention and RetentionSource types; GetDataLifecycleResponse now exposes GlobalRetention
  • Inference and _mvt content-type alignment: chat_completion_unified / stream_completion send Content-Type: application/json; _mvt uses the versioned application/vnd.elasticsearch+vnd.mapbox-vector-tile Accept header

Full Changelog: 8.19.21...8.19.22

What's Changed

• Regenerate client by @flobernd in #8897
  • Set explicit ContentType: application/json for ChatCompletionUnified and StreamCompletion streaming inference requests
  • Switch SearchMvt Accept header to application/vnd.elasticsearch+vnd.mapbox-vector-tile (Elasticsearch product-flavored content-type)
  • Bump Elastic.Transport to 0.17.1
  • Document default max_page_search_size values for pivot (500) and latest (5000) transforms

Full Changelog: 8.19.20...8.19.21