* __fix(dlq): release event handle when DlqObject has Event and is sent to S3 DLQ (#7244) (#7260)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Thu, 5 Nov 2026 13:45:33 -0800 EAD -> refs/heads/2.17, tag: refs/tags/2.17.0, refs/remotes/origin/2.17 When both a failure pipeline and S3 DLQ are configured, the DlqObject carries the full Event but may be released directly after S3 DLQ write instead of being sent to the failure pipeline. Previously, releaseEventHandle() silently dropped the release when event was not null, causing acknowledgement timeouts and kafka reprocessing. Release via event.getEventHandle().release() when event is present. (cherry picked from commit 03b33817bc94c718d17a6830197ff959156f4d01) Signed-off-by: Taylor Gray <tylgry@amazon.com> Signed-off-by: opensearch-ci-bot <opensearch-infra@amazon.com> Co-authored-by: Taylor Gray <tylgry@amazon.com> * __Bypass className resolution in Glue Schema Registry JSON deserialization (#7256) (#7259)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Thu, 5 Nov 2026 13:35:08 -0800 Wrap GlueSchemaRegistryKafkaDeserializer in SafeGlueSchemaRegistryKafkaDeserializer that bypasses Class.forName() for JSON schemas by parsing raw payload directly as a Map. For non-JSON schemas (AVRO, Protobuf), the delegate deserializer is used. (cherry picked from commit 66b9a481e6fd92d02064904ff533b2203d8c0840) Signed-off-by: Harsha Siva Sai Sannapaneni <sivasais@amazon.com> Signed-off-by: opensearch-ci-bot <opensearch-infra@amazon.com> Co-authored-by: Harsha Siva Sai Sannapaneni <harshasivasai85@gmail.com> Co-authored-by: Harsha Siva Sai Sannapaneni <sivasais@amazon.com> * __fix: Pass PostgreSQL database name as a driver property (#7255) (#7257)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Thu, 5 Nov 2026 12:55:22 -0800 (cherry picked from commit 83a70828ea5753fcd070ad2034b81ea22a1f7d45) (cherry picked from commit 2c6af61c00b735a01c324fec579377fa341663c1) Signed-off-by: Divyansh Bokadia <dbokadia@amazon.com> Signed-off-by: opensearch-ci-bot <opensearch-infra@amazon.com> Co-authored-by: Divyansh Bokadia <dbokadia@amazon.com> * __Adds the release notes for Data Prepper 2.17.0. (#7253) (#7258)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Thu, 5 Nov 2026 12:23:43 -0800 (cherry picked from commit 17ecb42d77c8dbd711179afd0848225bb2e2c6c3) Signed-off-by: David Venable <dlv@amazon.com> Signed-off-by: opensearch-ci-bot <opensearch-infra@amazon.com> Co-authored-by: David Venable <dlv@amazon.com> * __Prepare release 2.17.0 (#7245)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Wed, 4 Nov 2026 06:45:58 -0800 Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: dlvenable <293424+dlvenable@users.noreply.github.com> * __Support a scan delay for empty S3 folder partitions on the RDS source (#7238)__ [Siqi Ding](mailto:dingdd@amazon.com) - Mon, 2 Nov 2026 14:37:27 -0600 When a folder partition returns no objects, the S3 scan worker gives it up immediately, so a quiet source re-lists its folders back to back at S3 latency. A source with a trickle of changes therefore makes tens of ListObjectsV2 calls per second per worker even though almost every listing is empty. Add folder_partitions.empty_folder_rescan_delay to the S3 source. When set, an empty folder partition keeps its lease for the delay instead of being given up, so no worker lists it again until the delay has passed. The partition progress state is preserved, so idle folder deletion still works. If the lease cannot be extended, the partition is given up as before. Expose it on the RDS source as empty_scan_delay and pass it through the RDS template. The default is zero, which keeps the current behavior. Signed-off-by: Siqi Ding <dingdd@amazon.com> * __Increase the maximum JSON string length in the OpenSearch source and sink to 64 MB (#7222)__ [Siqi Ding](mailto:dingdd@amazon.com) - Mon, 2 Nov 2026 10:42:55 -0800 The opensearch source deserializes search responses using JacksonJsonpMapper instances created with Jackson's default StreamReadConstraints, which limit string values to 20,000,000 characters. Documents with a larger string field fail with StreamConstraintsException before an Event is created, and the source retries the index indefinitely. Raise the limit to JsonFactories.MAX_JSON_STRING_LENGTH (64 MB, the limit used by JacksonEvent since #7146) on the OpenSearch (RestClient and AwsSdk2) and Elasticsearch client mappers of the source. The opensearch sink also re-parses documents with default limits in a few places, so apply the same limit there: - BulkOperationFactory, used for update/upsert with document filters - FailedBulkOperationConverter, which otherwise writes an empty document to the DLQ - PullIngestionEnvelopeBuilder - PreSerializedJsonpMapper's JSON provider, used to parse responses Signed-off-by: Siqi Ding <dingdd@amazon.com> * __Update dependency setuptools to v83 (#7009)__ [mend-for-github-com[bot]](mailto:50673670+mend-for-github-com[bot]@users.noreply.github.com) - Mon, 2 Nov 2026 09:40:43 -0800 Co-authored-by: mend-for-github-com[bot] <50673670+mend-for-github-com[bot]@users.noreply.github.com> * __Update dependency werkzeug to v3.1.9 (#7176)__ [mend-for-github-com[bot]](mailto:50673670+mend-for-github-com[bot]@users.noreply.github.com) - Mon, 2 Nov 2026 09:40:20 -0800 Co-authored-by: mend-for-github-com[bot] <50673670+mend-for-github-com[bot]@users.noreply.github.com> * __fix(cloudwatch-logs): stop acknowledging undelivered logs (#7212)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Mon, 2 Nov 2026 09:38:20 -0800 * fix(cloudwatch-logs): stop acknowledging undelivered logs With acknowledgements enabled and no DLQ, the sink queued batches on an unbounded executor and returned immediately. It then retried failed batches forever and released every failure as delivered, so data was lost silently and memory grew during an outage. doOutput now hands batches to a fixed pool of `workers` threads and waits until every batch is settled. Each batch is retried up to max_retries, then sent to the DLQ or negatively acknowledged so the source can redrive it. The change also adds metrics for events released without delivery, sent to the DLQ, or negatively acknowledged. It also removes CloudWatchLogsLimits and Buffer methods that nothing calls any more, and trims unit tests that only repeated what another test already proves. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Retry until delivered when no DLQ or acks exist End-to-end acknowledgements are off by default, and a DLQ is optional. In that configuration a negative acknowledgement has nowhere to go, so a batch that ran out of retries was dropped after about 2.3 minutes of outage at the default max_retries. Such a batch is now retried until it is delivered or the sink shuts down. doOutput waits on it, so an outage applies backpressure to the source instead of dropping events, and memory stays bounded. max_retries still bounds retries when a DLQ is configured or every event has an acknowledgement set. Permanent errors (InvalidParameterException, HTTP 413) end the request after one attempt in both modes, so a batch that no retry can fix does not stall the pipeline. The backoff exponent is capped so unbounded retries cannot overflow the delay. Trade-off: while CloudWatch Logs is unavailable, the pipeline worker is blocked, so other sinks in the same pipeline stall with it. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Fail fast on missing log groups and fix review issues - Treat ResourceNotFoundException as permanent when neither create_log_group nor create_log_stream is set, so one missing destination no longer stalls the pipeline. - Add a permanentFailure reason so InvalidParameterException, HTTP 413 and missing resources are no longer counted as retriesExhausted. - Retry log group and stream creation on the next ResourceNotFoundException if the previous creation failed. - Run the undeliverable, unexpected and refused terminal steps even if takeReady throws, so oversized events keep their undeliverable handling. - Log once when max_retries does not apply because there is no DLQ or acknowledgement set. - Remove an internal product reference from the README. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Mark hot-path logs noisy and close retry gaps - Mark per-event, per-batch and per-attempt logs NOISY. - Stop catching StackOverflowError, as no other sink does. - Fail fast on a missing log group when only create_log_stream is set. - Test that send() retries until delivered only without a DLQ or acknowledgement set. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> --------- Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __Support Amazon OpenSearch Serverless NextGen time series indexes (#7225)__ [Divyansh Bokadia](mailto:dbokadia@amazon.com) - Fri, 30 Oct 2026 16:05:09 -0700 When aws.serverless is true, the opensearch sink detects NextGen time series indexes by reading the index.timestamp_field index setting. The result is cached per index. A failed lookup is not cached and is retried after the next bulk flush, since the flush may create the index. For create and index actions on a time series index, the sink adds the timestamp field to the document it sends without modifying the event: it keeps the field if present, otherwise moves @timestamp to it, otherwise sets it from the event time. When a bulk item is rejected with "document is missing required timestamp field [X]", the sink adds X to that document and retries it in the same flush, and records X for the index. This covers indexes that do not exist yet or whose settings cannot be read. All other errors keep the existing retry and DLQ behavior. Resolves #7221 Signed-off-by: Divyansh Bokadia <dbokadia@amazon.com> * __fix(rds-source): decrement numOfWorkers once per data file when acknowledgments are enabled (#7209)__ [photoruction-gb](mailto:80666489+photoruction-gb@users.noreply.github.com) - Fri, 30 Oct 2026 13:54:43 -0500 With acknowledgments enabled, DataFileScheduler decremented numOfWorkers twice per data file: once in runLoader.whenComplete and again in completeDataLoader, which the acknowledgment callback calls. The counter went negative, so the scheduler kept acquiring data file partitions while the only loader thread was busy. Those partitions waited in the executor queue without lease renewal, expired, and failed on the final save with ConditionalCheckFailed (Expected version: 1), causing the files to be reprocessed. Move the decrement out of completeDataLoader into the no-acknowledgment whenComplete path, so each data file releases its slot once, when its loader finishes. Type runLoader as CompletableFuture<Void>. Add DataFileSchedulerTest cases with acknowledgments enabled: the worker count returns to 0 after an acknowledged file, and no further partition is acquired while the loader is busy. Signed-off-by: Guillaume Bodi <guillaume.bodi@photoruction.com> * __Add GenAI enrichment settings to otel_traces: on/off switch, root propagation and coverage cache size (#7210)__ [Shenoy Pratik](mailto:pshenoy36@gmail.com) - Fri, 30 Oct 2026 06:57:20 -0700 * Add genai_root_propagation and genai_coverage_cache_max_size to otel_traces Root span enrichment copied agent data and token totals onto every trace root, including the HTTP or RPC span an agent runs behind. Now that agent runs are marked with derived.gen_ai.is_root, those copies are redundant and misleading. The new genai_root_propagation setting chooses which roots get them: ai_root (default) copies only onto a GenAI root or a root marked at the source, all_roots keeps the previous behavior, none copies nothing. The GenAI coverage cache size, previously fixed at 500,000 span ids, is now configurable with genai_coverage_cache_max_size. Rewrites the README GenAI section to document normalization, sub-key stripping, the agent root marker, the coverage cache and its sizing, root propagation modes, the token counting rule and the genAiCoverageCount metric. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Add genai_enrichment_enabled and address review comments Adds genai_enrichment_enabled (default true) to turn GenAI enrichment off entirely. When off, the processor skips normalization, sub-key stripping, marking and root enrichment, and creates no coverage cache. Review fixes: * genai_coverage_cache_max_size must be at least 1. * New genAiCoverageSpanIdCount gauge reports the cache's size in span ids, the unit of its limit; genAiCoverageCount counts traces. * An unknown genai_root_propagation value now fails with the allowed values instead of falling back to the default. * README: the cache size is a long, documents the new setting, metric and an example config. Adds config, enum and coverage unit tests, and processor-level tests for each propagation mode and for enrichment turned off. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Keep test dependencies unchanged and make the span id gauge current Checks the @Min(1) constraint on genai_coverage_cache_max_size through reflection instead of adding a validator test dependency. Applies pending cache writes before reading the weighted size, which Caffeine updates asynchronously, so genAiCoverageSpanIdCount reports the current count. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Rename the GenAI switch to disable_genai_enrichment and validate GenAI settings * genai_enrichment_enabled (default true) becomes disable_genai_enrichment (default false). * genai_root_propagation and genai_coverage_cache_max_size cannot be set when disable_genai_enrichment is true. Both are now unset by default and their getters return the defaults, so the check can tell whether the user set them. * genai_root_propagation values must match exactly; they are no longer lowercased. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> --------- Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> Co-authored-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * __Preserve standard span serialization when routing copies (#7198)__ [Karsten Schnitter](mailto:k.schnitter@sap.com) - Fri, 30 Oct 2026 14:03:48 +0100 * Preserve standard span serialization when routing copies Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * Copy standard span metadata independently Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> --------- Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * __Fix duplicate OTel attribute keys dropping entire OTLP request (#7155)__ [Karsten Schnitter](mailto:k.schnitter@sap.com) - Thu, 29 Oct 2026 21:00:40 +0100 Collectors.toMap without a merge function throws IllegalStateException when any OTLP attribute list carries repeated attribute keys. The exception propagated to the gRPC service layer, which discarded the whole batch. Fixed by replacing Collectors.toMap with a mutable-reduction collect(HashMap::new, (m, i) -> m.put(...), Map::putAll) across all attribute-converting methods in OTelProtoOpensearchCodec and OTelProtoStandardCodec, covering trace spans, resources, links, events, log records, metric data points, and exemplars. Last value wins for duplicates, matching the receiver behaviour described in the OTLP spec. The repeated pattern is extracted into a private collectAttributes helper in each codec. Resolves #7151 Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * __Fix duplicate document metric classification (#7207)__ [Karsten Schnitter](mailto:k.schnitter@sap.com) - Thu, 29 Oct 2026 20:36:18 +0100 Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * __Normalize DocumentDB source s3_prefix so a trailing slash does not change the staging layout (#7215)__ [Dinu John](mailto:dinujohn@amazon.com) - Tue, 27 Oct 2026 15:46:29 -0500 * Normalize DocumentDB source s3_prefix so a trailing slash does not change the staging layout A trailing slash in s3_prefix (for example "prefix/") was concatenated with the source's own delimiter, producing keys under "prefix//...". The extra empty path segment shifts the folder depth that the S3 read stage partitions on, so all staged objects were read as a single folder: slowly, and in key order rather than time order. Strip leading and trailing delimiters and collapse repeated delimiters so "prefix", "prefix/" and "/prefix/" produce the same layout. Signed-off-by: Dinu John <dinujohn@amazon.com> * Use a generic prefix in DocumentDBServiceTest Signed-off-by: Dinu John <dinujohn@amazon.com> * Add full license header to DocumentDBServiceTest Signed-off-by: Dinu John <dinujohn@amazon.com> --------- Signed-off-by: Dinu John <dinujohn@amazon.com> * __Fix giveUpPartition not executing in continuous scan mode (#7214)__ [Taylor Gray](mailto:tylgry@amazon.com) - Tue, 27 Oct 2026 12:39:52 -0500 The giveUpPartition call in the continuous scan mode path of ScanObjectWorker.processFolderPartition passed 0 as maxRetries. In LeaseBasedSourceCoordinator.giveUpPartition, the retry loop condition is 'attempt <= maxRetries', so with 0 the loop body never executes and the partition is never actually released. This caused folder partitions to remain stuck in ASSIGNED status until the 10-minute ownership lease expired, adding significant latency to continuous scan mode object detection-to-processing. Change maxRetries from 0 to MAX_RETRIES (5) to match the non-continuous scan mode path and the error handler. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Fix DocumentDB source stream never starting after a long export (#7213)__ [Dinu John](mailto:86094133+dinujoh@users.noreply.github.com) - Tue, 27 Oct 2026 11:42:32 -0500 The change stream waits for the STREAM global state, which is created once the export finishes. On large, long-running exports it could be skipped for good, so the pipeline stopped at "Initial load not complete": - ExportScheduler: the write that marks the total partition count as final is retried when it loses a version race with an export worker. The export partition is already completed at that point and is never acquired again, so a dropped write left the flag false forever. If every partition is already loaded when the flag is written, the scheduler starts the stream. - ExportWorker: a data query partition is completed before it is counted, so the late acknowledgment of an owner whose lease expired is not counted a second time. The stream starts once loaded >= total instead of only on exact equality, which an earlier double count could step over. Signed-off-by: Dinu John <dinujohn@amazon.com> * __Mark the outermost GenAI span of traces with a non-GenAI root (#7192)__ [Shenoy Pratik](mailto:pshenoy36@gmail.com) - Wed, 21 Oct 2026 19:14:30 -0700 * Mark the outermost GenAI span of traces with a non-GenAI root An agent served behind an HTTP handler has an HTTP root span, and the GenAI spans (invoke_agent, chat, execute_tool) are its descendants. Consumers that list agent traces by their root, such as OpenSearch Dashboards Agent Traces, miss these traces (#7191). When a trace's root is not a GenAI span, otel_traces now sets derived.gen_ai.is_root = true on each outermost GenAI span (a span with gen_ai.operation.name and no GenAI ancestor) and aggregates its descendants' tokens onto it. Spans flushed without their root are still marked. A marker already set by an SDK or by hand is kept, and the processor then marks nothing else in the trace. Root propagation of gen_ai.agent.name, gen_ai.request.model and the other keys now takes values from the outermost GenAI spans first, instead of whichever child came first in the batch. The README no longer says gen_ai.operation.name is propagated. Resolves #7191 Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Count each span's token usage once; cover agents calling agents Root enrichment summed token usage over every child, and the agent root marker did the same over its descendants. Agent spans such as invoke_agent often report the total of their own LLM calls, so those calls were counted twice; on a real Strands trace the root showed about twice the actual usage. Both now count only spans that have no ancestor reporting usage. Adds tests for an agent that calls another agent over HTTP in the same trace: only the calling agent is marked, and the called agent is marked only when its spans are flushed without the caller's. The README documents both. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Mark GenAI roots too; don't mark spans that arrive after their ancestors - The agent root marker now also goes on a trace root that is a GenAI span, so every agent trace carries one marker, wherever the agent sits. - Spans that arrive after their trace's root was processed are enriched on their own, without their ancestors, so a late chat span or the spans of a called service looked outermost and were marked. The processor now remembers, per trace and for trace_group_cache_ttl, which span ids are covered by a GenAI span or a marker (bounded by span id count), and looks there when a parent is not in the batch. - A span marked at the source also covers the spans below it. - README documents the remaining edge cases. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Address review: atomic per-trace coverage, one index per trace, safer values - Coverage now outlives the orphan flush (2 x trace_flush_interval + trace_group_cache_ttl), so spans flushed without their root still see what earlier batches covered; its size is reported as genAiCoverageCount. - The covered read, the marking and the write run atomically per trace (Caffeine compute), so pipeline workers handling spans of one trace at the same time cannot both mark an outermost span. - Each trace is indexed once: attributes are read once per span, traces without GenAI data are skipped, and one ancestor walk serves marking, coverage and token counting. - Start times compare as instants, not text (fractional precision). - An agent span reporting only output tokens keeps its value; the root check is symmetric too. - Token counts and propagated values are read tolerantly: numeric strings count, anything else is ignored instead of failing the batch. - Propagation uses the coverage too, like marking. - Only derived.gen_ai.is_root: true is a source marker; a span set to false is never marked. A debug log notes traces skipped for a marker. - README: the coverage window, the marker values, and the token rule's assumption that an ancestor's usage includes its subtree. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> --------- Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> Co-authored-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * __Map Traceloop session ids to gen_ai.conversation.id (#7173)__ [Shenoy Pratik](mailto:pshenoy36@gmail.com) - Wed, 21 Oct 2026 18:55:09 -0700 * Map Traceloop session ids to gen_ai.conversation.id Map traceloop.association.properties.{session_id,conversation_id,thread_id} to gen_ai.conversation.id in the openllmetry profile, so OpenLLMetry apps that set a session through association properties get the OTel GenAI attribute. The id is normalized on the span that carries it and is not copied to the root span: it is the key that groups traces into sessions, and the OTel GenAI semantic conventions define no root-span or propagation rule for it (they ask instrumentations to set it where it is readily available, including execute_tool spans since semantic-conventions-genai#518). Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Pick the first-listed GenAI mapping source; drop guessed Traceloop keys When several source attributes on one span mapped to the same gen_ai.* target, normalizeAttributes checked a copy of the attributes taken before the loop, so every later source overwrote the earlier one and the winner depended on attribute order. It now picks one source per target first, by position in the mappings file, then writes. Drops traceloop.association.properties.conversation_id and thread_id: Traceloop does not set them as association properties. Its set_conversation_id() and LangGraph thread_id handling write gen_ai.conversation.id directly. session_id stays: it is Traceloop's standard AssociationProperty.SESSION_ID. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * Test GenAI mapping priority: file order, duplicates, skipped entries Covers how buildLookupTable assigns priorities (file order across profiles, a duplicate source keeping its first target and priority, malformed entries taking no priority), and that the shipped OpenInference session.id outranks the Traceloop session_id mapping. Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> --------- Signed-off-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> Co-authored-by: Shenoy Pratik Gurudatt <4348487+ps48@users.noreply.github.com> * __fix(cloudwatch-logs): drop entity when a resolved value is blank or too long (#7203)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Wed, 21 Oct 2026 18:33:24 -0500 In dynamic-entity mode, a templated entity value that resolved to empty or over 512 characters made PutLogEvents reject the whole batch with a 400, which the sink retried indefinitely. Such events are now sent without an entity and counted in cloudWatchLogsEntityDropped. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __fix(source-crawler): renew lease during ack wait (#7202)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Wed, 21 Oct 2026 15:14:14 -0500 Worker partitions hold a 10 minute lease that was never renewed, while their acknowledgement set waits up to 2 hours. A sink stalled for more than 10 minutes let the lease expire, so the partition was re-acquired and its events re-emitted every 10 minutes, leaking an acknowledgement set each time. Renew the lease every 2 minutes from the acknowledgement set's progress check, as s3-source does, through a new renewPartitionOwnership on EnhancedSourceCoordinator. A blocked worker keeps its partition and a dead one releases it after one lease. Writes for a partition are serialized, and after a failed renewal the worker stops writing it and lets the lease lapse, so a stale write cannot overwrite a new owner. The ack callback also counted a success before completePartition ran, so a failed completion was counted as a success and its exception lost. Count a success only after completion, and count failures in a new acknowledgementCallbackFailures metric. Resolves #7200 Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __fix(kafka): pause and call poll() immediately on writeAll failure (#7196)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 19 Oct 2026 13:43:14 -0500 When the Kafka consumer buffer writeAll call fails (buffer full or circuit breaker), the consumer now pauses immediately on the first failure and calls doPoll() on every retry iteration. This keeps the member in the group by driving the network I/O that reads JoinGroup/SyncGroup responses and allows the heartbeat thread to function during rebalances. Previously, the consumer would retry writeAll for maxRetriesOnException iterations (calculated as max_poll_interval / 2000) with a 1s sleep between each retry before pausing. During a rebalance, the heartbeat thread disables itself after JoinGroup and only re-enables when poll() processes the JoinGroup response. Without poll() being called, the coordinator's JoinGroup response sits unread in the socket, and after session.timeout.ms (60s) expires with no heartbeat or SyncGroup, the coordinator removes the member. The main thread then gets UNKNOWN_MEMBER_ID on its next poll, triggering another rebalance. With this fix, poll() is called on every retry, so JoinGroup/SyncGroup responses are processed promptly and the heartbeat thread stays active. Rebalances complete in seconds even when the buffer is full. Also adds doPoll() in the pauseConsumePredicate path so that external pause conditions (like circuit breaker) also maintain group membership. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Make in-memory buffer and batch size configurable for kafka buffer, lower default values (#7194)__ [Taylor Gray](mailto:tylgry@amazon.com) - Sun, 18 Oct 2026 11:59:27 -0500 Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Add trace-analytics-plain-raw-v2 index type (#7188)__ [Suresh N S](mailto:41610499+nssuresh2007@users.noreply.github.com) - Fri, 16 Oct 2026 09:14:38 -0700 * Add trace-analytics-plain-raw-v2 index type Add an opensearch sink index_type, trace-analytics-plain-raw-v2, that writes the same span documents as trace-analytics-plain-raw to a new otel-v2-apm-span alias, following the otel-v2-apm-service-map pattern. It has its own template (V1 and index-template formats) and its own ISM policy, otel-v2-apm-span-policy, whose ISM template matches otel-v2-apm-span-*. The template is the plain span template plus: - keyword mappings for string fields under the nested events.attributes and links.attributes, for clusters that do not support text fields inside nested fields; - index.append_only.enabled: false, so that spans can be written with the document ID ${traceId}/${spanId} on clusters where new indices are append-only by default. This setting requires OpenSearch 3.0. Signed-off-by: Suresh N S <nssuresh@amazon.com> * Compose the v2 span template from component templates Build the trace-analytics-plain-raw-v2 index template from component templates instead of a full copy of the plain span template, so the span mappings are defined in one file only. The index template now only sets index.append_only.enabled: false and composes, in order: - otel-v2-apm-span-plain: the template block and version of the existing plain span template (otel-v1-apm-span-index-standard- template.json), built at startup; - otel-v2-apm-span-nested-keyword: keyword mappings for strings under events.attributes and links.attributes. The sink creates or updates these component templates before the index template, each only when its version is higher than the one on the cluster, using the same check as for index templates. Component templates configured through component_templates are still appended after them. Because a V1 template cannot compose component templates, this index type always uses template_type index-template and is rejected with distribution_version es6. The legacy V1-format copy of the v2 template is removed. Signed-off-by: Suresh N S <nssuresh@amazon.com> * Add full license header to component template wrapper The license header check requires all five header lines from CONTRIBUTING.md; the new wrapper and its test only had the first two. Signed-off-by: Suresh N S <nssuresh@amazon.com> --------- Signed-off-by: Suresh N S <nssuresh@amazon.com> * __fix(kafka): apply sts_header_overrides to MSK IAM SASL auth (#7187)__ [hozkaya2000](mailto:63765239+hozkaya2000@users.noreply.github.com) - Tue, 13 Oct 2026 10:09:31 -0700 Route role-based AWS_MSK_IAM SASL (data-plane) authentication through a dedicated callback handler that sources credentials exclusively from Data Prepper's header-aware STS provider, so aws.sts_header_overrides (x-amz-source-arn / x-amz-source-account) are applied on the data-plane sts:AssumeRole used for broker auth. The handler implements AuthenticateCallbackHandler directly (never extends IAMClientCallbackHandler), so the vendored library never assembles its own header-less MSKCredentialProvider. The ROLE path is now unconditional: credentials always flow through our provider regardless of whether overrides are configured. Resolves #6923 Signed-off-by: Halit Ozkaya <hozkaya@amazon.com> * __Upgrade OpenTelemetry proto Java bindings and add event_name support for log records (#7148)__ [Karsten Schnitter](mailto:k.schnitter@sap.com) - Tue, 13 Oct 2026 15:47:23 +0200 * Upgrade OpenTelemetry proto Java bindings to 1.11.0-alpha Updates io.opentelemetry.proto:opentelemetry-proto from 1.3.2-alpha to 1.11.0-alpha, bringing support for all OTLP protocol changes through proto v1.11.0 including the stabilised event_name field in LogRecord, EntityRef in Resource definitions, and the new ProcessContext message. Resolves #1269 Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * Add support for event_name field in OTel log records Maps the top-level LogRecord.event_name field (introduced in OTel proto v1.5.0) to a new eventName property on OpenTelemetryLog. Both the standard and OpenSearch codecs now decode the field from incoming requests, and the standard codec encoder writes it back when re-serialising logs to protobuf. Resolves #6505 Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * Generate OTel proto Java classes from source to keep protobuf-java at 3.25.5 opentelemetry-proto:1.11.0-alpha requires protobuf-java:4.35.1, which conflicts with the project-wide pin of protobuf-java:3.25.5. Instead of consuming the pre-built Maven JAR, generate the Java classes and gRPC service stubs directly from the .proto files bundled in the opentelemetry-proto:1.11.0-alpha:sources@jar artifact using protoc:3.25.5 and protoc-gen-grpc-java:1.70.0. This keeps protobuf-java firmly at 3.25.5 across all modules. Changes: - otel-proto-common: add com.google.protobuf Gradle plugin v0.9.4, switch to java-library, generate proto and gRPC classes from sources JAR, expose protobuf-util/grpc-stub/grpc-protobuf as api dependencies - Remove direct opentelemetry-proto dependency from all other modules (otel-trace-source, otel-logs-source, otel-metrics-source, otlp-source, otel-trace-raw-processor, otel-metrics-raw-processor, http-common, otlp-sink, kafka-plugins, aggregate-processor, service-map-stateful, all e2e-test projects) - service-map-stateful: replace direct proto artifact with project dep on otel-proto-common, removing the now-unnecessary protobuf-java exclude - Remove opentelemetry version catalog entry from settings.gradle - Exclude generated proto sources from Spotless and Checkstyle - Fix implicit Spotless/antlr task dependency in data-prepper-expression Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * Fix e2e test compilation: add otel-proto-common dependency The e2e trace and log tests import io.opentelemetry.proto.* classes directly. After switching otel-proto-common to generate those classes from source, the e2e projects need an explicit dependency on otel-proto-common to get them on the integration test compile classpath. Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * Fix Gradle 8.8 implicit task dependency in data-prepper-logstash-configuration Gradle 8.8 detects that spotlessJava reads from the ANTLR output directory produced by generateGrammarSource without a declared dependency, causing a build failure when the two tasks run in parallel. Add an explicit dependsOn to satisfy Gradle's validation and restrict the Spotless target to src/ so the generated files are not formatted. Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> --------- Signed-off-by: Karsten Schnitter <k.schnitter@sap.com> * __fix(opensearch-sink): S3 DLQ takes precedence over failure pipeline (#7189)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 12 Oct 2026 19:27:20 -0500 Previously, when both a per-sink S3 DLQ and a global failure pipeline (dlq_pipeline) were configured, the failure pipeline always took precedence. This meant the S3 DLQ configured on the OpenSearch sink was never used. This caused problems when the failure pipeline writes to the same OpenSearch cluster that is experiencing issues — the failure pipeline would also fail, blocking acknowledgements indefinitely (with acks enabled) or causing negative acks upstream. New priority order for handling failed documents: 1. S3 DLQ file writer (dlqFileWriter) — if configured 2. S3 DLQ writer (dlqWriter) — if configured on the sink 3. Failure pipeline (dlq_pipeline) — fallback if no sink-level DLQ 4. Log warning and release(false) — nothing configured If S3 DLQ write fails, fall back to the failure pipeline. If no failure pipeline exists either, release with false. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __fix(core): Skip circuit breaker for DLQ pipeline buffer (#7185)__ [Taylor Gray](mailto:tylgry@amazon.com) - Thu, 8 Oct 2026 13:58:36 -0500 The DLQ (failure) pipeline uses a HeadlessPipelineSource which writes events to its buffer via sendEvents(). When a circuit breaker is active, the buffer is wrapped with CircuitBreakingBuffer which throws TimeoutException when the breaker is open. This causes DLQ writes to block indefinitely (with acks enabled) or release with false (without acks), preventing acknowledgements from completing upstream. Since the DLQ pipeline exists specifically to drain failed events, blocking it under memory pressure defeats its purpose and can cascade into negative acknowledgements on the source pipeline's Kafka buffer. Skip applying CircuitBreakingBuffer to HeadlessPipelineSource pipelines, matching the existing skip for PipelineConnector sources. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __fix(prometheus-sink): Retry on unprocessed requests (GOAWAY, connection reset) (#7182)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 5 Oct 2026 23:11:35 -0500 Add retry handling for UnprocessedRequestException in the Prometheus sink's Armeria WebClient. Previously, only HTTP status codes 429, 502, 503, and 504 triggered retries. Connection-level failures like HTTP/2 GOAWAY frames, connection resets, and connection refused errors were not retried, causing the request to fail immediately and trigger negative acknowledgements upstream in the pipeline. HTTP/2 GOAWAY is a normal connection lifecycle event sent by load balancers and server proxies during connection rotation. The request was never processed by the server, so it is safe to retry on a new connection. Changes: - Add .onUnprocessed() to the Armeria RetryRuleWithContent chain, which matches UnprocessedRequestException (covers GoAwayReceivedException, connection reset, connection refused) - Extract shared Backoff instance to avoid duplication - Add unit tests for GoAwayReceivedException handling Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Support continuous scan mode on RDS source (#7181)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 5 Oct 2026 16:10:31 -0500 Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Support disabling the ttl cleanup on documentdb source with s3 source folder partitioning (#7172)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 5 Oct 2026 12:18:55 -0500 Signed-off-by: Taylor Gray <tylgry@amazon.com> * __[aws-lambda] Add retryable_status_codes option to retry transient HTTP status codes (#7165)__ [Luis Norman](mailto:lanorman14@gmail.com) - Fri, 2 Oct 2026 13:06:53 -0700 Signed-off-by: LuisNorman <lunorman@amazon.com> Co-authored-by: LuisNorman <lunorman@amazon.com> * __Make DimensionalTimeSliceCrawler partition creation wait configurable (#7179)__ [lksfrnz](mailto:lksfrnz@amazon.com) - Thu, 1 Oct 2026 20:58:33 -0500 Adds CrawlerSourceConfig.getPartitionCreationWait() (default 5 minutes). DimensionalTimeSliceCrawler now receives the source config through an @Inject constructor and uses that wait instead of the hard-coded 5 minutes, so a source changes the wait by overriding the getter. The wait must be between 0 and 60 minutes and is validated when the crawler is created. The existing (CrawlerClient, PluginMetrics) constructor keeps the 5-minute default. Signed-off-by: Lukas Franz <lksfrnz@amazon.com> * __Emit typed external-dependency nodes and RED metrics in APM service map processor (#7147)__ [Adam Tackett](mailto:105462877+TackAdam@users.noreply.github.com) - Thu, 1 Oct 2026 11:40:22 -0700 * feat: synthesize typed external-dependency nodes and RED metrics in APM service map processor Signed-off-by: Adam Tackett <tackadam@amazon.com> * feat: emit per-type dependency identity attributes on synthesized nodes Signed-off-by: Adam Tackett <tackadam@amazon.com> * feat: suppress unresolved/raw-IP dependency peers; add tests and docs Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): address review — exclude dep attributes from node identity, gate by span kind, inherit env, harden URL parsing, hoist providers Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): address bot re-review - keep AWS::*/ARN dependency names, strict IP-literal filter, safer URL path parsing, exact span-kind match Signed-off-by: Adam Tackett <tackadam@amazon.com> * chore: retrigger CI (unrelated flaky failures) Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): gate dependency nodes behind opt-in flag, cap cardinality, name DBs by instance, fix double counts Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): add .keyword sub-field to dependencyAttributes template (v2), document host-less DB nodes and query paths, lock in DB name fallback Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): admit capped dependencies by call count, keep overflow out of the operation budget, make the overflow metrics per-call counters, log derivation failures Signed-off-by: Adam Tackett <tackadam@amazon.com> * chore(apm): use Javadoc for new fields; set service map template version to 1 Signed-off-by: Adam Tackett <tackadam@amazon.com> * feat(apm): enable dependency_nodes by default Signed-off-by: Adam Tackett <tackadam@amazon.com> * fix(apm): sticky dependency admission across windows, per-type overflow names, spanKind label on messaging series; document rollout, naming spec and data contract Signed-off-by: Adam Tackett <tackadam@amazon.com> --------- Signed-off-by: Adam Tackett <tackadam@amazon.com> Co-authored-by: Adam Tackett <tackadam@amazon.com> * __Remove stale tableId mapping in RDS source when a table id is reused (#7162)__ [gg-mskm](mailto:masakide0514gg@gmail.com) - Wed, 30 Sep 2026 15:38:22 -0500 The MySQL RDS source keeps an in-memory `tableId -> TableMetadata` map that is built from binlog TABLE_MAP events. `handleTableMapEvent` returned early for any table that is not of interest without touching the map, so an entry recorded for a table of interest survived indefinitely. The binlog `table_id` is assigned by the server from its table definition cache and is not stable: it can be reassigned to a different table across a server restart, and also when a table is evicted from and reloaded into the cache. When a table id previously bound to a table of interest was reassigned to a table that is not of interest, `isValidTableId` still accepted the row events for that table because the stale entry was present and pointed at a table of interest. Those rows were then decoded using the wrong column definitions, which silently wrote corrupted documents to the sink and raised type-conversion errors for the rows whose types did not line up. Remove the mapping held for a table id whenever a TABLE_MAP event shows the id now refers to a table that is not of interest, so subsequent row events for that id are skipped as intended. Resolves #7125 Signed-off-by: Masaki Mori <masakigg@amazon.co.jp> Co-authored-by: Masaki Mori <masakigg@amazon.co.jp> * __Consolidates the code for configuring 64 MB when deserializing from JSON. (#7167)__ [David Venable](mailto:dlv@amazon.com) - Tue, 29 Sep 2026 12:24:42 -0700 Adds a new JsonFactories class that creates a default JsonFactory. Use this in code where we updated to 64 MB already and in the ndjson and json codecs. Signed-off-by: David Venable <dlv@amazon.com> * __Add component_templates option to the opensearch sink (#7161)__ [Vinay Krishna Pudyodu](mailto:vinkrish.neo@gmail.com) - Mon, 28 Sep 2026 17:37:24 -0700 Adds the configured component template names to the composed_of list of the index template the sink creates, including the built-in index_type templates. This lets users layer index settings (e.g. index.codec, or a parquet-backed data format) onto built-in templates such as trace-analytics-plain-raw without copying them. Requires template_type index-template. Resolves #7142 Signed-off-by: Vinay Krishna Pudyodu <vinkrish.neo@gmail.com> * __Now that we support 64 MB strings in Events we also need to support this larger value on the event_json codec. (#7166)__ [David Venable](mailto:dlv@amazon.com) - Mon, 28 Sep 2026 17:04:39 -0700 Updates the EventJsonInputCodec to support a 64 MB maximum string length. Signed-off-by: David Venable <dlv@amazon.com> * __feat(cloudwatch-logs): Send dynamic per-request headers (#7157)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Wed, 23 Sep 2026 15:27:08 -0500 The sink could only attach headers once at client build, so every PutLogEvents request carried the same values. Templated values in header_overrides are now resolved per event and attached to their own request, letting a single sink label each request with headers derived from the events it carries. Header values join the existing entity key attributes in the grouping key, so events that differ only by a header are sent as separate requests while events that share a key are batched into one. Literal header values keep the original client-level behavior. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __fix(cloudwatch-logs): Count UTF-8 bytes (#7154)__ [James Liu](mailto:71547730+Jiaaming@users.noreply.github.com) - Tue, 22 Sep 2026 15:31:38 -0500 Measure event size using the exact UTF-8 bytes buffered for delivery. This prevents request and event size checks from undercounting non-ASCII log data. Signed-off-by: James Liu <jamesliu.jiaming@gmail.com> * __test(mongodb): fix flaky ExportSchedulerTest await anchors (#7111)__ [kiran536](mailto:veeravelly.kirankumar@gmail.com) - Sat, 19 Sep 2026 05:40:42 -0700 Three tests in ExportSchedulerTest awaited an intermediate coordinator interaction and then called future.cancel(true), which can interrupt ExportScheduler.run() before it makes the remaining calls the assertions expect. Observed in CI as: TooFewActualInvocations: coordinator.getPartition(...) Wanted 3 times: ExportSchedulerTest.java:173 But was 2 times: ExportScheduler.run(ExportScheduler.java:79) ExportScheduler.run(ExportScheduler.java:83) Line 172, the getPartition inside markTotalPartitionsAsComplete, never ran because cancel landed after createPartition but before it. Anchor each await on the last coordinator interaction of the scheduler's iteration, saveProgressStateForPartition(globalState, ...), which every other assertion in these tests happens at or before. Contributes to #3481. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> Co-authored-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * __Fix flaky EndToEndRawSpanTest in peer forwarder e2e tests (#6981)__ [Sotaro Hikita](mailto:70102274+lawofcycles@users.noreply.github.com) - Sat, 19 Sep 2026 05:38:43 -0700 * Fix flaky EndToEndRawSpanTest in peer forwarder e2e tests Raise trace_flush_interval from 5 to 30 in the raw-span e2e pipeline configs so a peer-forwarded root span is not garbage-collected away from its buffered children, and report all spans with missing fields via soft assertions instead of throwing on the first one. Ref #6721 Signed-off-by: Sotaro Hikita <bering1814@gmail.com> * Use a dedicated pipeline config for the peer forwarder e2e test rawSpanEndToEndTest runs two nodes without peer forwarding and relies on the 5 second trace_flush_interval to release buffered child spans for otel_trace_group backfill, so raising the shared value broke it. Keep raw-span-e2e-pipeline.yml at 5 and give rawSpanPeerForwarderEndToEndTest its own config with trace_flush_interval 30. Ref #6721 Signed-off-by: Sotaro Hikita <bering1814@gmail.com> * Add license header to new e2e pipeline config Signed-off-by: Sotaro Hikita <bering1814@gmail.com> --------- Signed-off-by: Sotaro Hikita <bering1814@gmail.com> * __Add Splunk HEC source plugin (RFC #6823) (#7105)__ [Srikanth Padakanti](mailto:srikanth_padakanti@apple.com) - Thu, 17 Sep 2026 16:51:52 -0500 * Add a Splunk HTTP Event Collector (HEC) source plugin Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * Add a getDefaultSsl override hook to the HTTP server configuration Sources which should be secure by default can override getDefaultSsl to enable ssl when the pipeline author does not set it, mirroring the existing getDefaultPort and getDefaultPath hooks. The interface method defaults to false so that every existing HTTP source is unchanged. Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * Use BackgroundThreadFactory for the Splunk HEC ack cleanup thread Replaces the hand-rolled ThreadFactory with the shared BackgroundThreadFactory for consistency with the rest of the project. Also switches the expired ack cleanup from compute to computeIfPresent. The two are equivalent here, but compute required a null-state guard that no test could reach, which held the module below its required JaCoCo coverage minimum of 1.0. Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * Build Splunk HEC events with EventFactory and share the ingest path Extracts every event construction concern into HecEventBuilder, which builds events through the injected EventFactory instead of JacksonEvent.builder. Metadata resolution, flattening, field merging and timestamp parsing now live in a plain class that HecEventBuilderTest covers directly with 32 tests, without standing up an HTTP port. Collapses processEventWithAck, processEventWithoutAck and processRawRequest into a single handleIngestRequest entry point plus a shared writeRecords. Authentication, data channel validation, request metrics and the buffer write now run on one path that no endpoint can bypass, and /ack authenticates before reporting whether acknowledgements are enabled. This takes SplunkHecService from 745 to 389 lines. Two inconsistencies surface from sharing the path: /raw now sets the same sourcetype, source and host metadata attributes that /event sets, and /raw now applies the per-token default fields it previously ignored. A failed acknowledgement also removes the pending ack rather than leaving it pending until it expires. Configuration changes: * ssl now defaults to true, so the certificate and key are required unless ssl is explicitly disabled. * Configuring the shared authentication plugin is rejected, since the HEC protocol authenticates with the tokens option and two concurrent authentication mechanisms would be ambiguous. * Renames ack_expiry to acknowledgement_expiry. * Removes use_forwarded_headers, which nothing read. * Documents why default_sourcetype is a free-form string rather than an enum: Splunk sourcetypes are defined per deployment with no fixed set. Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * Update the Splunk HEC source README for the configuration changes Documents the acknowledgement_expiry rename, the removal of use_forwarded_headers, the secure-by-default ssl behavior, why the authentication plugin is not supported alongside HEC tokens, why default_sourcetype is a free-form string, and that both endpoints now attach the same routing metadata. Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> --------- Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * __Increase the maximum string length in an Event to 64 MB. (#7146)__ [David Venable](mailto:dlv@amazon.com) - Wed, 16 Sep 2026 12:10:07 -0700 Resolves #7145. Signed-off-by: David Venable <dlv@amazon.com> * __Parse arrays in parse_json and parse_ion. (#7141)__ [David Venable](mailto:dlv@amazon.com) - Tue, 15 Sep 2026 18:35:25 -0700 * Parse arrays in parse_json and parse_ion. The parse_json and parse_ion processors currently do not support parsing arrays that are not enclosed in an object. This change has them parse the arrays if the `destination` is set. It is not possible to parse into the root, so the parse will fail if no destination field is set when an array is detected. Additionally, this implementation also allows parsing of scalar values. Resolves #5705 Signed-off-by: David Venable <dlv@amazon.com> * Fix the pointer path bug found by the PR auto-review. Signed-off-by: David Venable <dlv@amazon.com> --------- Signed-off-by: David Venable <dlv@amazon.com> * __Add fields parameter to Jira /rest/api/3/search/jql request (#7144)__ [Harsha Siva Sai Sannapaneni](mailto:harshasivasai85@gmail.com) - Tue, 15 Sep 2026 14:28:18 -0700 The migration from /rest/api/3/search to /rest/api/3/search/jql (#7106) did not account for a behavior difference: the old endpoint returned all navigable fields by default, whereas the new endpoint returns only issue IDs. Without the fields parameter, IssueBean.fields is null, causing a NullPointerException in JiraItemInfo.getKeyAttributes() on every crawl cycle. Fix: add fields=project,created,updated to the search/jql request. These are the only fields consumed from the search response by JiraItemInfo. Resolves #7143 Signed-off-by: Harsha Siva Sai Sannapaneni <sivasais@amazon.com> Co-authored-by: Harsha Siva Sai Sannapaneni <sivasais@amazon.com> * __Fix NPE when a cascading-action column is updated to NULL in the rds source (#7137)__ [David Venable](mailto:dlv@amazon.com) - Fri, 11 Sep 2026 08:37:29 -0700 Currently, the rds source throws a NullPointerException and drops the entire change event when an UPDATE sets a column to NULL on a table that is a parent in a foreign key relationship with a cascading action (CASCADE, SET_NULL, or SET_DEFAULT). This change builds the changed-columns map with a HashMap instead of Collectors.toMap(), which rejects null values. It also adds unit tests covering an UPDATE that sets a cascading column to NULL. Resolves #7129 Signed-off-by: David Venable <dlv@amazon.com> * __Creates a new configuration for acknowledgements and makes the shutdown timeout configurable. (#7127)__ [David Venable](mailto:dlv@amazon.com) - Tue, 8 Sep 2026 12:11:03 -0700 This adds a new acknowledgements section in data-prepper-config.yaml and includes a shutdown_timeout which is a configurable Duration. It retains the same default value of five seconds as is currently in use. Signed-off-by: David Venable <dlv@amazon.com> * __feat(key-value): add destination_conflict_strategy for per-field merge in key-value processor (#7101)__ [lfpantin](mailto:lmayaud@amazon.com) - Tue, 1 Sep 2026 13:52:33 -0700 feat(key-value): add destination_conflict_strategy to key-value processor The key-value setting `overwrite_if_destination_exists` was all-or-nothing when a destination was set: it replaced the whole destination map or skipped the write entirely. This commit adds `destination_conflict_strategy` with skip, overwrite, merge_preserve_existing_keys, and merge_overwrite_existing_keys. The merge strategies resolve conflicts per-field on an existing destination map. overwrite_if_destination_exists is deprecated but still honored (false -> skip, true -> overwrite); setting both options is rejected. The default stays overwrite, so existing pipelines are unaffected. Write logic is extracted into a reusable EventDestinationWriter, with unit, processor, and integration tests plus a README update. Resolves #7091 Signed-off-by: Luis Pantin <lmayaud@amazon.com> * __feat(osgi): add opt-in Apache Felix plugin framework (#6917)__ [kiran536](mailto:veeravelly.kirankumar@gmail.com) - Tue, 1 Sep 2026 08:31:16 -0700 * feat(osgi): add opt-in Apache Felix plugin framework Add an opt-in plugin-loading mode backed by an embedded Apache Felix OSGi framework, enabled with -Ddata-prepper.plugin.framework=osgi. When the property is unset, Data Prepper uses the existing classpath loader. In that legacy mode plugin discovery is functionally identical to before: the same ServiceLoader providers are found in the same order, the OSGi framework never starts, no OSGi code path executes, and only a couple of inert Spring beans are constructed. Plugins are prepared for OSGi at build time rather than adapted at runtime. A new Gradle plugin module data-prepper-gradle-plugins/osgi-plugin (plugin id org.opensearch.dataprepper.plugin) bakes an OSGi manifest into the plugin JAR at build time using bnd (biz.aQute.bnd). It is consumed through includeBuild in pluginManagement and is not published to any repository yet, so for now only projects in this repository can apply it; in-repo plugins opt in by applying the plugin id. It does not apply maven-publish, since no publication or repository is configured for it. There is no runtime adaptation and no BundleAdapter. At startup in OSGi mode StaticBundleLoader installs the pre-built bundle JARs directly, resolves and starts them with fail-fast behavior, a startup summary log, and Micrometer metrics. A JAR that lacks an OSGi manifest (no Bundle-SymbolicName) fails fast with an actionable message pointing the author to the Gradle plugin. The plugin-framework-osgi module embeds the Felix lifecycle (FelixPluginManager), bootstraps it in OSGi mode through a Spring-managed runner (OsgiFrameworkRunner), and exposes a PluginProvider backed by the OSGi service registry (OsgiPluginRegistry). Baked plugins declare their @DataPrepperPlugin packages through the DataPrepper-Plugin-Classes manifest header, which the Gradle plugin generates from META-INF/data-prepper.plugins.properties, and LegacyPluginBundleActivator registers those classes as OSGi services. Data Prepper API packages are exported to bundles with semver version attributes. The exported package set is generated at build time by a Gradle task that writes META-INF/osgi-shared-packages.properties listing every org.opensearch.dataprepper.* package except org.opensearch.dataprepper.plugins.*, so there is no hand-maintained list. The version comes from the existing data-prepper-api DataPrepperVersion, formatted into the OSGi three-part string by a small package-private toOsgiVersion helper; the separate DataPrepperApiVersion class was removed. This is consistent with the data-prepper-api backward-compatibility contract from #6607, so a plugin built on 2.15 resolves on a 2.16 host. The shared set also includes jakarta.validation, jakarta.validation.constraints, and com.fasterxml.jackson.annotation (annotations only: jackson-databind and jackson-core are intentionally not exported, preserving isolation), plus org.slf4j and the framework package org.opensearch.dataprepper.plugin.osgi so bundles can load LegacyPluginBundleActivator from the system bundle. Felix framework storage resolves under System.getProperty("data-prepper.dir")/data/osgi/felix-cache, matching the MaxMindConfig idiom, with a java.io.tmpdir fallback when the property is unset such as in tests. It does not use the system temp directory by default. The directory name is fixed rather than process-unique so that FRAMEWORK_STORAGE_CLEAN_ONFIRSTINIT actually reclaims it on the next start; a per-process name would orphan one cache directory per restart and grow without bound on a long-lived host or a container with a persistent data volume. A fixed name is safe because a single framework instance owns the directory for the life of the process, and a cache left by an unclean shutdown is cleaned on the next first init. Caches orphaned by the earlier process-suffixed naming are purged at startup. Only the tmpdir test fallback keeps a PID suffix, where concurrent test JVMs would otherwise share one directory. The OSGi service registry is the single source of truth for plugin discovery. LegacyPluginBundleActivator loads each bundle's annotated classes once at bundle start and publishes one service per plugin class, carrying every name the plugin answers to — its name, deprecated name, and alternate names — as a multi-valued service property. OsgiPluginRegistry reads those properties, selected with an LDAP filter on dataprepper.plugin.name, and resolves the named classes through the publishing bundle rather than re-scanning bundles for annotations, so no plugin class is loaded twice and no unrelated service is force-activated through getService(). Fragment bundles are attached at resolve time and cannot be started, so they are excluded from the started-bundle list and from the bundlesLoaded counter rather than inflating them; the startup summary still lists them, reported by their actual state. BundleHealthCheck reports a bundle whose classloader cannot be verified as not isolated, agreeing with its own null-wiring path, instead of swallowing the error and passing. OSGi metrics are recorded against the application's CompositeMeterRegistry, injected into OsgiFrameworkRunner, rather than Metrics.globalRegistry, which Data Prepper does not bind to the registries it exports. The dependency is an Optional because this bean is found by package scan and several contexts scan the OSGi package without the metrics configuration; when no registry bean exists, metrics go to a detached registry rather than to the global one. The Gradle plugin emits bnd Export-Package directives per clause, since a bnd directive applies only to the clause it is attached to; attaching -noimport once to a joined package list would leave every package but the last importing its own exports. Each declared package is trimmed so a list written with spaces after the commas does not produce clauses with leading whitespace. bnd is pinned to 6.4.0: 7.0.0 is compiled to Java 17 bytecode, and because the bnd plugin is declared in settings.gradle every build evaluates it, so a Java 11 Gradle JVM cannot load BndBuilderPlugin at all and Java 11 jobs such as verify-api-compatibility fail before any task runs. BundleResolutionErrorTranslator turns Felix resolution failures into readable diagnostics. BundleHealthCheck verifies framework, bundle, and classloader isolation. BundleClassLoaderScope manages the thread context classloader at the plugin invocation boundary so ServiceLoader and SPI lookups resolve against the bundle classloader. PluginHealthProbe is a seam for future functional health checks. The OsgiFrameworkRunner bean is always instantiated by Spring but returns early in @PostConstruct unless data-prepper.plugin.framework=osgi. The Felix framework JAR ships on the runtime classpath as a small (<1MB) library with zero transitive dependencies and is not exercised in legacy mode. Scope is limited to classloader isolation and deploy-time validation. Bundles install once at startup and stop at shutdown, the same lifecycle as classpath plugins. There is no production hot-reload; PluginHotLoader is test-scoped only. Data Prepper-owned system properties use the data-prepper. namespace: data-prepper.plugin.framework and data-prepper.plugin.bundles.dir. The genuine OSGi and Felix properties (org.osgi.*, felix.*) are unchanged. Testing: an integration test OsgiIT in a new integrationTest source set under plugin-framework-osgi builds a DefaultPluginFactory like DefaultPluginFactoryIT, loads a test plugin through it, and proves the load happened through OSGi rather than the classpath. It asserts that FrameworkUtil.getBundle is non-null, the plugin classloader is a BundleReference, the classloader is not the application classloader, and the application classloader throws ClassNotFoundException for the plugin class because the plugin is supplied only as a bundle JAR and never on the application classpath. Unit tests drive OsgiPluginRegistry entirely from mocked service references, covering resolution of every published name, the plain-string property shape, and services that are incomplete or whose classes cannot be loaded. StaticBundleLoader is tested with a real fragment attached to a real host to assert the fragment is absent from the started list and does not increment the bundlesLoaded counter. Each behavior above is covered by a test that fails if that behavior is reverted: metrics are asserted to land in the injected registry and to be absent from Metrics.globalRegistry; OsgiPluginRegistry is asserted never to call getService(); LegacyPluginBundleActivator is asserted to publish all of a plugin's names as one multi-valued String[] property; BundleHealthCheck is asserted to report a bundle as not isolated both when its wiring is null and when reading its classloader throws; PluginProviderLoader is asserted to return registered providers ahead of classpath providers; DefaultPluginFactory is asserted to resolve a provider registered after its construction and to name both empty-provider cases in its error message; and toPackageGlobs is asserted to trim a package list written with spaces after the commas. The Gradle plugin has unit and functional (GradleRunner) tests, including a two-package project that proves the per-clause Export-Package fix: with the directive on only the last clause, bnd emits a self-import for the first package, and the test asserts that import is absent. Wiring changes: settings.gradle adds the plugin-framework-osgi module, the new data-prepper-gradle-plugins/osgi-plugin (consumed via includeBuild in pluginManagement), the data-prepper-test/osgi-test-plugin subproject, a Felix version-catalog entry, and the bnd plugin. build-resources.gradle adds plugin-framework-osgi to coreProjects. data-prepper-core declares a runtimeOnly dependency on the new module. PluginProviderLoader is made public with a registerProvider() hook and caches the framework-mode flag at construction. DefaultPluginFactory re-reads the providers on every lookup rather than snapshotting them in its constructor, which is what makes the OSGi provider visible: the OSGi framework registers it from a @PostConstruct and there is no Spring ordering edge forcing that to run before the first plugin lookup. Both ends document that a snapshot must not be reintroduced without adding such an edge, and the empty-provider error message now distinguishes a missing PluginProvider service file from an OSGi framework that has not finished starting. Tests verify these changes, and legacy behavior is unchanged. In OSGi mode getPluginProviders() returns the OSGi-registered providers ahead of the classpath providers and DefaultPluginFactory takes the first match, so a bundle shadows a classpath plugin of the same name. This is intended and is now documented in the module README. The example and test plugin is data-prepper-test/osgi-test-plugin, not a released data-prepper-plugins module. It demonstrates a minimal @DataPrepperPlugin built as an OSGi bundle through the new Gradle plugin. Resolves #6760 Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * fix(osgi): address review follow-ups and minimize new dependency surface Follow-up to the OSGi plugin framework commit, covering the remaining reviewer feedback and the High-severity findings the PR Code Analyzer reported. Kept as a separate commit so the reviewed history is not rewritten. Reviewer follow-ups: SPI resolution now covers plugin construction. PluginCreator switches the thread context classloader to pluginClass.getClassLoader() around constructor.newInstance and restores it in a finally block, and only when it differs from the current TCCL. That makes it a strict no-op for the classpath framework, where plugin classes come from the application classloader, while under OSGi it is the plugin's own bundle classloader. It lives in data-prepper-plugin-framework, needs no OSGi types, and covers extensions too because ExtensionLoader shares the PluginCreator bean. Runtime invocation is deliberately not wrapped: that would put a TCCL swap on the per-batch hot path. The README now lists exactly which three boundaries are covered and states plainly what is not. Provider registration order is now guaranteed rather than merely early. OsgiFrameworkRunner implements the new PluginProviderRegistrar interface and DefaultPluginFactory takes it as Optional, so Spring must finish the runner's @PostConstruct — which starts Felix and registers the OSGi provider — before the factory can be constructed, and no plugin can be looked up before the factory exists. Optional because the only implementation lives in a module core depends on at runtime only; in legacy mode the argument is simply empty. The per-lookup provider re-read stays as defense in depth rather than as the ordering guarantee. The bundles-directory scan no longer rejects build by-products. StaticBundleLoader skips -sources.jar and -javadoc.jar, which Gradle and Maven publish next to a real artifact and which are never bundles. Every other JAR without an OSGi manifest is still rejected, so a plugin that forgot to apply the Gradle plugin fails loudly at startup instead of going missing at pipeline start. Fixed a stale-bundle defect that was masking test results. stageOsgiBundles was a Gradle Copy, and Copy never deletes, so a bundle JAR built under an earlier project version stayed in the staging directory beside the current one. OsgiIT points Felix at that whole directory, OSGi installs both versions of the same Bundle-SymbolicName, both register the same plugin name, and which one wins a lookup is undefined. The task is now a Sync, and OsgiIT asserts that exactly one bundle with the test plugin's symbolic name is installed so the same class of problem cannot return silently. Dependency surface: Removed the third-party bnd build plugin declaration from settings.gradle. bnd is now reached only through the included osgi-plugin build's own runtime classpath, so this repository declares no third-party build plugin; only the first-party includeBuild remains. Dropped the bnd plugin and its generated manifest from plugin-framework-osgi itself, along with the validateBundle check that asserted the header. That manifest was dead weight: data-prepper-core declares runtimeOnly project(':plugin-framework-osgi'), so this module is the OSGi host on the application classpath and is never installed as a bundle. Bundles reach org.opensearch.dataprepper.plugin.osgi through FRAMEWORK_SYSTEMPACKAGES_EXTRA, not through a bundle manifest. Removed three dependencies after verifying each was unnecessary. org.osgi:osgi.core is redundant because the OSGi core API, including the FrameworkUtil and BundleReference the integration tests use, is contained in the Felix framework JAR. logback-classic and micrometer-test were unused. Duplicated javax.inject and javax.annotation-api declarations in the integrationTest block were inherited already. That leaves exactly two new third-party coordinates, both pinned exactly rather than by range: Apache Felix (org.apache.felix.framework 7.0.5, Apache-2.0), which is the framework itself and is inert unless -Ddata-prepper.plugin.framework=osgi is set, and bnd (biz.aQute.bnd.gradle 6.4.0, Apache-2.0 OR EPL-2.0), which is build-time only and confined to the included build. bnd stays on the latest 6.x because 7.x is compiled to Java 17 bytecode and this repository builds on a Java 11 Gradle JVM. Every other dependency of the new module is already used at the same version elsewhere in Data Prepper. Provenance and license for both are recorded in plugin-framework-osgi/README.md. Hardened the stale bundle-cache purge in FelixPluginManager. It matched any directory whose name merely began with the cache name, so an unrelated directory such as felix-cache-backup was eligible for deletion, and it used a symlink-following directory check, so a link planted under the OSGi directory could redirect the walk outside it. A candidate is now removed only when it is a direct child of the OSGi directory, its name matches felix-cache-<digits> exactly, and it is a real directory rather than a symbolic link. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * test(mongodb): fix pre-existing flaky ExportSchedulerTest await anchors Three tests in ExportSchedulerTest awaited an intermediate coordinator interaction and then called future.cancel(true), which can interrupt ExportScheduler.run() before it makes the remaining calls the assertions expect. This surfaced on build (21) as: TooFewActualInvocations: coordinator.getPartition(...) Wanted 3 times: ExportSchedulerTest.java:173 But was 2 times: ExportScheduler.run(ExportScheduler.java:79) ExportScheduler.run(ExportScheduler.java:83) Line 172, the getPartition inside markTotalPartitionsAsComplete, never ran because cancel landed after createPartition but before it. Anchor each await on the last coordinator interaction of the scheduler's iteration, saveProgressStateForPartition(globalState, ...), which every other assertion in these tests happens at or before. The race predates this branch: the times(3) getPartition assertion was added in #4553, and data-prepper-plugins/mongodb is otherwise untouched by this PR. Tracked under #3481. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * test: fix flaky ProcessorValidationIT @SingleThread assertion ProcessorValidationIT.test_events_processed_when_SingleThread_processor intermittently failed in CI with "Expected: <1> but: was <0>" at verifySingleThreadUsage. The multi-processor-pipeline/SingleBatch case submits a single batch to a four-worker pipeline, so a worker can legitimately finish without ever reading a batch, leaving its @SingleThread processor instance unused. Asserting that every instance was used by exactly one thread therefore asserted something the pipeline does not guarantee. Assert the actual @SingleThread contract instead: no instance is used by more than one thread, no thread is shared across instances, and at least one instance was used. This still catches an instance being shared between workers, which the previous assertion could not distinguish from an idle worker. Also make SingleThreadEventsTrackingTestProcessor.PROCESSORS a CopyOnWriteArrayList. Instances are constructed on the process worker threads and the list is read from the test thread, so a plain ArrayList was unsafely published, unlike its sibling ConcurrentHashMap. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * fix(osgi): export third-party packages with versions and name unsatisfied imports Three defects made the OSGi path unusable for any real plugin. They only affect the opt-in `-Ddata-prepper.plugin.framework=osgi` path; the default plugin framework is untouched. 1. `third-party.packages` was generated without a `version` attribute. An OSGi export with no version defaults to `0.0.0`, which satisfies no version range, and bnd computes ranges for every third-party import it finds in a plugin JAR. `mutate-string-processors`, for example, imports `com.fasterxml.jackson.annotation;version="[2.21,3)"`, so it could never resolve. Each package is now exported with the version resolved for this build rather than a hardcoded one, so the two cannot drift apart. 2. `org.slf4j.spi` was never exported. SLF4J 2.x bytecode references the provider SPI, so bnd adds `org.slf4j.spi` to the imports of virtually any plugin that logs, and exporting `org.slf4j` alone is not enough. 3. `StaticBundleLoader` reported resolution failures before starting any bundle, so the message could only ever be "Bundle state is INSTALLED (not RESOLVED)" and never named the unsatisfied requirement. `FrameworkWiring.resolveBundles` returns only a boolean, so the detail is now recovered from the `BundleException` that `Bundle.start()` throws and passed to the existing `BundleResolutionErrorTranslator`. Fragments are skipped because they cannot be started. A bundle importing a package the host does not export now reports: Plugin '...' could not be resolved: it requires package com.fasterxml.jackson.databind;version=[2.21.0,3.0.0) but the host/other bundles do not provide a compatible version. Verified by bundling a real plugin (`mutate-string-processors`) with its JAR held out of `lib/`: it resolves, goes ACTIVE and processes an event, which was impossible before this change. `DataPrepperOsgiPackages.SLF4J_PACKAGE` still pins `org.slf4j` to a hardcoded version; that is left for a follow-up since the pinned version satisfies the range plugins ask for. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> --------- Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> Co-authored-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * __feat(os-s3-dlq): add codec option for JSON or NDJSON output (#7108)__ [Teja Reddy](mailto:tejareddy.bvsk@wbd.com) - Thu, 27 Aug 2026 07:04:38 -0700 feat(os-s3-dlq): add codec option for JSON or NDJSON output Signed-off-by: Teja Reddy <tejareddy.bvsk@wbd.com> * __Add JDBC-based source coordination store plugin (#6757)__ [Sotaro Hikita](mailto:70102274+lawofcycles@users.noreply.github.com) - Tue, 25 Aug 2026 12:06:17 -0700 Add JDBC-based source coordination store plugin Signed-off-by: Sotaro Hikita <bering1814@gmail.com> * __feat(prometheus-sink): add instance_label (#7084)__ [Teja Reddy](mailto:bvsktejareddy@gmail.com) - Tue, 25 Aug 2026 10:48:09 -0500 Time series labels are drawn only from metric attributes, so every Data Prepper instance running the same pipeline emits series with identical labels. Prometheus, AMP, Cortex, and Mimir deduplicate on series plus timestamp, so they keep one instance's sample and drop the rest without reporting an error. An N-instance deployment reports roughly 1/N of the true value and nothing surfaces the loss. instance_label names a label whose value the sink resolves at startup from this instance's IPv4 address, so each instance writes a distinct series and consumers recover the true value by aggregating the label away. The value is resolved rather than configured, so one configuration shared by every instance is enough. Managed deployments need that: OpenSearch Ingestion runs every OCU of a pipeline from a single definition, leaving nowhere to write a per-instance value. The address is not taken from the hostname. Hostname resolution depends on the hostname being present in the name service, and HostContext falls back to the shared constant "unknown" when it is not, which would give every instance the same value and reintroduce the collision. The sink instead asks the operating system which address it would use to reach the configured url, by connecting a datagram socket to the endpoint and reading the socket's local address. Connecting a datagram socket performs a routing lookup and sends nothing, so this opens no connection and costs one syscall at startup. On an instance holding more than one address this yields the address the endpoint observes rather than an arbitrary choice among them. Where the lookup cannot answer, such as an endpoint which does not resolve at startup, the sink falls back to the IPv4 addresses on the local interfaces, sorted so the choice is stable across restarts. NetworkInterface is already how peer forwarding decides whether an address belongs to this instance, and rds-source already derives a per-instance server id from the local address, so the address is an established identity for an instance. An instance which cannot resolve an address at all fails at startup instead. The label name must be a valid Prometheus label name, must not begin with "__", and must not be one of the names the sink adds to individual series ("le", "ge", "quantile"), which remote write would reject as a repeated label name. Signed-off-by: Teja Reddy <tejareddy.bvsk@wbd.com> * __Add support for rate limiting at a group level (#7103)__ [Divyansh Bokadia](mailto:dbokadia@amazon.com) - Mon, 24 Aug 2026 12:40:05 -0700 Signed-off-by: Divyansh Bokadia <dbokadia@amazon.com> * __feat(cloudwatch-logs): support dynamic per-entity grouping in sink (#7034)__ [Kirtan Kakadiya](mailto:35823164+KirtanKakadiya@users.noreply.github.com) - Mon, 24 Aug 2026 08:24:15 -0700 PutLogEvents accepts exactly one entity per request. When any entity key_attributes or attributes value contains a ${...} reference, the sink now partitions buffered events by their resolved key attributes and sends one request per distinct entity, so multi-source streams get correct per-resource entity attribution. Concurrently buffered entities are bounded by entity.max_cardinality; idle groups are released, and overflow falls back to a shared group with no entity. Static-entity mode is preserved and unchanged when no value is templated. Signed-off-by: Kirtan Kakadiya <kirtanhk@amazon.com> Co-authored-by: Kirtan Kakadiya <kirtanhk@amazon.com> * __Migrate Jira search from removed /rest/api/3/search to /rest/api/3/se… (#7106)__ [Harsha Siva Sai Sannapaneni](mailto:harshasivasai85@gmail.com) - Fri, 21 Aug 2026 16:57:55 -0700 Migrate Jira search from removed /rest/api/3/search to /rest/api/3/search/jql Closes #7104 Signed-off-by: Harsha Siva Sai Sannapaneni <sivasais@amazon.com> * __Add Kafka source observability metrics (#7102)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Thu, 20 Aug 2026 17:06:33 -0500 Add metrics to the Kafka source, published through the existing Micrometer registry, so operators can diagnose scaling and reliability issues that were previously visible only in logs, or not at all: - Scaling: per-partition lag, assigned partitions, active readers, and configured workers, so a bottleneck can be seen as partition-bound vs worker-bound. - Rebalance activity: rebalance and partitions-revoked counts. - Processing latency: poll to buffer handoff. - Federated auth (azure_federated): token-refresh success and failure counts (failures tagged by a bounded errorType), and headroom in seconds to the next proactive token refresh. The failure-handling counters from the issue (commit failures, buffer write failures, offset-reset failures) are intentionally left out here; they are covered by #7088. Because the azure_federated callback handler is created reflectively by kafka-clients, its metrics object is handed off through a singleton (KafkaSourceAuthMetricsProvider), mirroring AwsCredentialsSupplierProvider, and is registered only for azure_federated sources. Also make the azure_federated auth failure logging generic and actionable: log the underlying exception with its stack trace instead of asserting a single presumed cause, and drop the SENSITIVE marker from messages whose arguments are not credentials. Resolves #7074 Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __Bump 1password/load-secrets-action to v5.0.1 (data-prepper) (#7107)__ [Peter Zhu](mailto:zhujiaxi@amazon.com) - Thu, 20 Aug 2026 15:23:10 -0400 Signed-off-by: Peter Zhu <zhujiaxi@amazon.com> * __Fix CloudWatch sink DLQ to use default AWS credentials (#7076)__ [Reshmi Vijayan](mailto:reshmi.vijayan@ibm.com) - Tue, 18 Aug 2026 15:45:58 -0700 Fix CloudWatch sink DLQ to use default AWS credentials Signed-off-by: revijay <reshmi.vijayan@ibm.com> * __Fix JSON Pointer escape handling in JacksonEvent (#7079)__ [Reshmi Vijayan](mailto:reshmi.vijayan@ibm.com) - Tue, 18 Aug 2026 09:26:51 -0700 Fix JSON Pointer escape handling in JacksonEvent Signed-off-by: revijay <reshmi.vijayan@ibm.com> * __Stop the acknowledgement set monitor thread when stopping the Data Prepper container. (#7100)__ [David Venable](mailto:dlv@amazon.com) - Tue, 18 Aug 2026 07:01:38 -0700 Signed-off-by: David Venable <dlv@amazon.com> * __Updates AWS SDK to 2.53. (#7097)__ [David Venable](mailto:dlv@amazon.com) - Fri, 14 Aug 2026 10:16:52 -0700 Signed-off-by: David Venable <dlv@amazon.com> * __Support custom index names in otel_trace_group processor (#6936) (#6946)__ [Srikanth Padakanti](mailto:srikanth_padakanti@apple.com) - Thu, 13 Aug 2026 12:34:52 -0500 * Support custom index names in otel_trace_group processor (#6936) Adds an optional `indices` list to the otel_trace_group processor config. The processor previously hardcoded the search target to the TRACE_ANALYTICS_RAW alias ("otel-v1-apm-span"), so users whose raw spans were stored under custom index names or patterns could not use the processor. The new field accepts concrete index names, aliases, and patterns such as `otel-v1-apm-span-*`. Backward compatibility is preserved: when `indices` is omitted the processor uses the existing TRACE_ANALYTICS_RAW alias, so deployed pipelines.yaml files continue to work without changes. The field is marked @NotEmpty so explicit misconfiguration fails fast at startup with a clear validation error rather than silently broadcasting the search to all indices. Resolves #6936 Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * Use full Apache-2.0 license header on otel_trace_group config files OTelTraceGroupProcessorConfig.java and OTelTraceGroupProcessorConfigTest.java had no license header on main; this PR added the short 3-line form. Use the full 9-line header per CONTRIBUTING.md. Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> --------- Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * __fix(rss): harden feed poller error handling and config validation (#7085)__ [Tony Piazza](mailto:tony.piazza@aiven.io) - Thu, 13 Aug 2026 11:24:31 -0500 * fix(rss): harden feed poller error handling, validate backoff/tracker config, and cover auth + validation in tests Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * fix(rss): address review feedback on error handling, logging, and validation Signed-off-by: Tony Piazza <tony.piazza@aiven.io> --------- Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * __Update BouncyCastle to 1.85. (#7087)__ [David Venable](mailto:dlv@amazon.com) - Wed, 12 Aug 2026 10:39:48 -0700 The 1.85.2 patch exists only for some dependencies. To keep these in sync I'm updating only to 1.85. Signed-off-by: David Venable <dlv@amazon.com> * __Feature/aggregate sum action (#7080)__ [harshwadhawe](mailto:harsh.wadhawe@tamu.edu) - Tue, 11 Aug 2026 19:40:27 -0500 * __feat(rss): rewrite rss source to fix serious bugs and add support for multiple feeds (#7036)__ [Tony Piazza](mailto:tony@piazzaconsulting.com) - Tue, 11 Aug 2026 11:59:56 -0500 * feat(rss): rewrite rss source to fix serious bugs and add support for multiple feeds Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * test(rss): replace jsr310 dep with DataPrepperDurationDeserializer Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * address review: fix dedup edge cases, backoff scheduling, license headers, docs Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * address review: enforce min polling_frequency, stagger feed startup, fix license header Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * Rename pub_date to publication_date, guid to item_id, and validate feed names Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * address review: name pool threads, unify item_id with dedup key, add configurable request_timeout Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * chore: trigger CI Signed-off-by: Tony Piazza <tony.piazza@aiven.io> --------- Signed-off-by: Tony Piazza <tony.piazza@aiven.io> * __fix(rds-source): Use null-safe comparison for CDC update event column values (#7082)__ [Hai Yan](mailto:8153134+oeyh@users.noreply.github.com) - Tue, 11 Aug 2026 11:26:36 -0500 Signed-off-by: Hai Yan <oeyh@amazon.com> * __Add auto-detection of pre-compression JSON headers in decompression engines (#7025)__ [yavmanis](mailto:yavmanis@amazon.com) - Fri, 7 Aug 2026 10:40:35 -0700 Add auto-detection of pre-compression headers in decompression engines. Supports JSON-only detection for Snappy. Signed-off-by: Manisha Yadav <yavmanis@amazon.com> * __Add Kafka source consumer config properties (#7075)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Thu, 6 Aug 2026 05:49:31 -0700 Expose metadata_max_age and acknowledgments_expiry_reset on the source consumer, and fix SourceTopicConfig validation constraints. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __Add configurable connections_max_idle for topics (#7017)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Wed, 5 Aug 2026 09:37:56 -0700 * Add configurable connections_max_idle for topics Expose Kafka's connections.max.idle.ms as a per-topic connections_max_idle option on the Kafka source. This lets operators recycle idle connections before brokers that close them early. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Apply connections_max_idle to the Kafka buffer Expose getConnectionsMaxIdle() on the shared TopicConsumerConfig interface instead of a dedicated marker interface, so the Kafka buffer consumer picks up connections_max_idle in addition to the source. The factory reads it through a plain interface call, dropping the instanceof downcast. Also constrain the value to a minimum of one second via @DurationMin. Default remains 540000ms, so existing pipelines are unchanged. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> --------- Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __fix(rds-source): Use schema-qualified table name for join metadata enrichment in PostgreSQL (#7044)__ [Hai Yan](mailto:8153134+oeyh@users.noreply.github.com) - Tue, 4 Aug 2026 00:15:01 -0500 Signed-off-by: Hai Yan <oeyh@amazon.com> * __fix(dynamodb-source): Decrement numOfWorkers before giveUpPartition to prevent scheduler stall (#7070)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 3 Aug 2026 14:41:00 -0500 When a DataFileLoader completes with an exception and acknowledgments are enabled, the whenComplete callback previously called giveUpPartition before decrementing numOfWorkers. If giveUpPartition threw an unchecked exception, the decrement was skipped and numOfWorkers stayed at MAX_JOB_COUNT permanently, preventing the DataFileScheduler from acquiring any new partitions. This fix moves the decrement before giveUpPartition and wraps the giveUpPartition call in a try/catch. The counter is always decremented regardless of whether partition release succeeds, ensuring the scheduler can continue acquiring work. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Add gRPC keepalive max_connection_age settings on OTel sources (#6745) (#6947)__ [Srikanth Padakanti](mailto:srikanth_padakanti@apple.com) - Mon, 3 Aug 2026 07:59:29 -0700 Expose `max_connection_age` and `connection_drain_duration` on every Data Prepper source that exposes a gRPC server (otlp, otel-logs, otel-trace, otel-metrics). Both settings are optional `Duration` fields on the existing per-source configuration. When set, they are applied to the Armeria `ServerBuilder` via `maxConnectionAge(Duration)` and `connectionDrainDuration(Duration)`. When unset they preserve current behavior (no upper bound on connection age). This addresses the horizontal-scaling problem where long-lived gRPC channels never re-resolve DNS to discover new replicas. Configuring `max_connection_age` forces clients to reconnect periodically and pick up new endpoints. Implementation details: - New `Duration` fields with `@DurationMin` / `@DurationMax` validation on `OTLPSourceConfig`, `OTelLogsSourceConfig`, `OTelTraceSourceConfig`, `OTelMetricsSourceConfig`. - `ServerConfiguration` in `http-common` carries the two new fields so `otlp-source` and `otel-metrics-source` pick them up through the shared `CreateServer.createGRPCServer(...)` path. `otel-logs-source` and `otel-trace-source` apply them directly to their inline `ServerBuilder`. - `hibernate-validator` is promoted to the version catalog (`libs.hibernate.validator`) so all four source modules share a single pinned version, replacing the inline pin previously in `otlp-source`. Resolves #6745 Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * __Match OpenSearch shard routing in pull-ingestion IndexRouter (#7013)__ [Srikanth Padakanti](mailto:srikanth_padakanti@apple.com) - Mon, 3 Aug 2026 07:54:32 -0700 Match OpenSearch shard routing in pull-ingestion IndexRouter The pull-ingestion IndexRouter computed the Murmur3 hash over UTF-8 bytes, but OpenSearch's Murmur3HashFunction.hash(String) hashes UTF-16 little-endian bytes, so documents could be routed to a different partition than the shard OpenSearch places them on. The shard calculation also used floorMod(hash, numberOfShards) instead of OpenSearch's floorMod(hash, routingNumShards) / routingFactor, diverging whenever number_of_routing_shards differs from number_of_shards. Hash the UTF-16 little-endian bytes, source number_of_routing_shards from the index settings (falling back to number_of_shards), and scale by routingFactor so partition assignment matches OpenSearch's OperationRouting.generateShardId. Adds a cross-check against OpenSearch's authoritative Murmur3HashFunction known values and known-answer shard tests. Resolves #6993 Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * __feat(flatten): Add configurable flatten_separator parameter (#7042)__ [Taylor Gray](mailto:tylgry@amazon.com) - Wed, 29 Jul 2026 14:59:15 -0500 Add flatten_separator parameter to the flatten processor to allow configuring the character used to join nested key names. Previously hardcoded to ".", now supports any single character (e.g., "_", "-"). Default remains "." for backward compatibility. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Fix azure_federated Kafka auth after token refresh (#7040)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Wed, 29 Jul 2026 09:40:07 -0700 Fix azure_federated Kafka auth after token refresh The Kafka source stopped consuming about one hour after start and did not recover. getToken() cached one token object and returned that same instance. Kafka identifies OAUTHBEARER tokens on the SASL Subject by object identity: a re-login adds the new token and removes the old one. With the same instance, the add does nothing and the remove empties the Subject, so the consumer can no longer authenticate. Return a new token instance from each getToken() call, but keep the cached value and absolute expiry so the token exchange does not repeat and the refresh schedule does not change. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __Extend length() expression function to support Map, List, and Array types (#7031)__ [bhavsshr](mailto:bhavsshr@amazon.com) - Wed, 29 Jul 2026 08:45:33 -0700 Previously, length() only supported String values (returning character count). This enhancement adds support for: - Map: returns the number of keys (e.g., length(/metadata) returns 2 for {"a":1,"b":2}) - List: returns the number of elements (e.g., length(/tags) returns 3 for ["x","y","z"]) - Array: returns the number of elements (raw Java arrays) This enables users to write conditions based on collection sizes in expressions, e.g., 'length(/body/details) > 0' to check if a map/list is non-empty, or 'length(/records) > 100' to route based on batch size. The change is fully backward-compatible - existing String behavior is unchanged. Testing: - Added tests using @Nested classes for grouping (String, Map, List, Array, Error cases) - Used mocked Event for Map, List, Array, and Error tests to guarantee branch coverage - All code paths covered for 100% JaCoCo instruction coverage - getFunctionName(), all type branches, and all error paths tested Signed-off-by: bhavsshr <bhavsshr@amazon.com> * __Add integration test verifying processors work with empty, null, and bare key configurations (#7012)__ [Siqi Ding](mailto:dingdd@amazon.com) - Tue, 28 Jul 2026 13:52:37 -0700 Signed-off-by: Siqi Ding <dingdd@amazon.com> * __Reduce global permissions (#7035)__ [Peter Zhu](mailto:zhujiaxi@amazon.com) - Mon, 27 Jul 2026 17:57:14 -0400 Signed-off-by: Peter Zhu <zhujiaxi@amazon.com> * __Add code diff analyzer/reviewer to dataprepper repository (#7033)__ [Peter Zhu](mailto:zhujiaxi@amazon.com) - Mon, 27 Jul 2026 17:37:00 -0400 Signed-off-by: Peter Zhu <zhujiaxi@amazon.com> * __Detect zstd compressed objects in automatic compression mode (#6909)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Mon, 27 Jul 2026 12:42:06 -0700 CompressionOption.fromFileName() backs the "automatic" compression option for the S3 source (both SQS-notification and scan paths) and the S3 enrich processor. It only recognized ".gz" (GZIP) and ".snappy" (SNAPPY) and silently fell back to NONE for everything else, so zstd-compressed objects were passed to the codec without decompression and failed to parse — even though ZSTD is a valid CompressionOption with a working ZstdDecompressionEngine already wired in. Map the ".zst" (canonical, per the zstd CLI and RFC 8478) and ".zstd" (informal alternate spelling) extensions to CompressionOption.ZSTD so automatic detection decompresses these objects correctly. Behavior is unchanged unless a user explicitly configures compression: automatic. Add unit tests covering both extensions. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __Add configurable shard_discovery_interval to DynamoDB stream source (#7001)__ [Archit Kaushik](mailto:asarchit65@gmail.com) - Mon, 27 Jul 2026 11:26:21 -0700 * Add configurable shard_discovery_interval to DynamoDB stream source The DynamoDB stream source's LeaderScheduler ran shard discovery on a hardcoded 1-minute interval. Because DynamoDB Streams shards roll over continuously, newly opened child shards were not read until the next discovery pass, putting a ~1-minute floor on shard-rotation latency and surfacing as high EndToEndLatency while PipelineLatency stayed low. Add an optional shard_discovery_interval on the stream config, defaulting to 1 minute so behavior is unchanged unless set. A minimum of 1 second is enforced to avoid excessive DescribeStream/ListShards calls. Resolves #7000 Signed-off-by: Archit Kaushik <kausarch@amazon.com> * refactor: use @DurationMin annotation and drop pipeline-parser test dep Address review feedback: - Replace custom @AssertTrue validation with @DurationMin(seconds = 1) from hibernate-validator, matching the pattern used by aws-lambda, cloudwatch-logs, and geoip-processor plugins. - Remove testImplementation on :data-prepper-pipeline-parser; use jackson-datatype-jsr310 JavaTimeModule for Duration deserialization in tests (ISO-8601 format), matching the aws-lambda test convention. - Validate via jakarta Validator in tests instead of calling a method. Signed-off-by: Archit Kaushik <architkshk@gmail.com> Signed-off-by: Archit Kaushik <kausarch@amazon.com> * refactor: simplify LeaderScheduler construction and fix license header Address dlvenable's review feedback: - Extract lease interval into a local variable and always use the 5-arg LeaderScheduler constructor, removing the two-constructor branch. - Make DEFAULT_SHARD_DISCOVERY_INTERVAL public so DynamoDBService can reference it for the null-streamConfig fallback. - Add the full license header to StreamConfigTest.java per CONTRIBUTING.md. Signed-off-by: Archit Kaushik <architkshk@gmail.com> Signed-off-by: Archit Kaushik <kausarch@amazon.com> --------- Signed-off-by: Archit Kaushik <kausarch@amazon.com> Signed-off-by: Archit Kaushik <architkshk@gmail.com> Co-authored-by: Archit Kaushik <kausarch@amazon.com> * __Add azure_federated authentication to Kafka source (#7011)__ [Nikhil Bagmar](mailto:40037072+bagmarnikhil@users.noreply.github.com) - Thu, 23 Jul 2026 12:24:11 -0700 * Add azure_federated authentication to Kafka source Enable the Kafka source to consume from Azure Event Hubs by federating an AWS identity to Azure Active Directory for OAuth, so no Azure client secret is stored in the pipeline. Credentials use Data Prepper's standard AWS configuration, so one pipeline works across AWS credential setups. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Close STS client and share JAAS constants in azure_federated Kafka auth - AzureFederatedTokenProvider now exposes close() (lock-guarded) and AzureFederatedCallbackHandler.close() delegates to it, releasing the lazily-created StsClient on shutdown instead of leaking it. - Promote the OPT_* JAAS option keys on AzureFederatedCallbackHandler to public and reference them from KafkaSecurityConfigurer so the writer and reader share one definition and can't drift. - Replace the repeated "SASL_SSL" literals in KafkaSecurityConfigurer with a SASL_SSL_PROTOCOL constant. - Add unit tests for close() delegation and StsClient cleanup. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * Simplify azure_federated Kafka auth configuration Drop the redundant azure_ prefix from config keys since they already sit under azure_federated. Remove tenant_id, which duplicated the tenant already present in token_endpoint. Resolve the AWS region from the data-prepper-config.yaml default when the pipeline config omits it, rather than requiring aws.region. Remove the orphaned integration-test gradle properties whose test is not in the repo. Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> --------- Signed-off-by: Nikhil Bagmar <nikhilbagmar73@gmail.com> * __fix(s3-source): Log full stack trace in S3ObjectWorker buffer write error (#7015)__ [Taylor Gray](mailto:tylgry@amazon.com) - Tue, 21 Jul 2026 14:43:34 -0500 The existing error log only prints e.getMessage() which is null for exceptions like NullPointerException, making the error impossible to diagnose. Log the full exception with stack trace instead. Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Fix flaky file-source unit tests with an injectable Clock and controllable executor (#6984) (#6995)__ [Srikanth Padakanti](mailto:srikanth_padakanti@apple.com) - Tue, 21 Jul 2026 09:16:04 -0700 FileReader.readLoop and its backpressure/batch logic branched on System.currentTimeMillis(), so the max-read-time and drain-timeout tests depended on wall-clock timing. On a slow or loaded runner the loop could break before the first channel.read, leaving Mockito stubs unused and failing with UnnecessaryStubbingException. Inject a java.time.Clock into FileReader (public constructor defaults to Clock.systemDefaultZone(), package-private constructor accepts a Clock), matching the existing pattern in ServiceMapStatefulProcessor, OTelApmServiceMapProcessor, and PipelineShutdown. Tests now drive time with a mocked Clock: the read stub advances the fake clock past the deadline after delivering data once, so exactly one read happens deterministically. Removes all Thread.sleep and wall-clock assumptions from these tests. FileReaderPoolTest bookkeeping tests submitted readers to a real thread pool where the reader completed on a background thread and mutated the active/pending counts before the assertions ran. Add createPoolWithNonRunningExecutor so queueing, promotion, and close logic run synchronously. Tests that verify execution keep the real executor and await via verify(..., timeout(...)). Signed-off-by: Srikanth Padakanti <srikanth_padakanti@apple.com> * __feat(split-event): Add split_when conditional parameter (#7014)__ [Taylor Gray](mailto:tylgry@amazon.com) - Mon, 20 Jul 2026 16:59:24 -0500 Add split_when parameter to the split_event processor, consistent with the conditional expression pattern used by other processors (e.g., delete_when, add_when, parse_when). When split_when evaluates to false, the event passes through unchanged. Closes #6992 Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Log data stream config notices once per stream instead of per document (#6979)__ [nsa-softpay](mailto:nsa@softpay.io) - Mon, 20 Jul 2026 09:16:07 -0700 The opensearch sink logs configuration notices for data streams (first-write-wins document_id, unsupported routing, and action coercion to `create`) inside determineAction, which runs once per event. On high-volume pipelines this emits one WARN per ingested document, dominating log output. These notices describe configured, expected behavior rather than a per-record fault, so dedupe them by (data stream, notice type) and log each at most once per data stream. The returned action is unchanged. Resolves #6972 Signed-off-by: nsa-softpay <nsa@softpay.io> * __Return HTTP 400 with informative message for client-side protobuf validation errors (#6989)__ [Indraja Babu](mailto:156961807+indrajababu@users.noreply.github.com) - Sat, 18 Jul 2026 10:41:43 -0700 When a client sends a protobuf request containing invalid UTF-8 in string fields (e.g., InstrumentationScope name/version), the pipeline currently returns HTTP 500 with a generic error. This change returns HTTP 400 with a client-actionable message explaining the issue. Changes: - Override apply() in GrpcRequestExceptionHandler to intercept InvalidProtocolBufferException before Armeria's default short-circuit - Detect via stable cause-chain type check (not brittle message string) - Apply same fix to HTTP path handlers in otel-logs-source and otel-trace-source - Add wire-level integration test proving HTTP 400 for invalid UTF-8 protobuf - Add test for unsupported grpc-encoding header rejection Signed-off-by: Indraja Babu <indbabu@amazon.com> * __fix(sqs): Change acknowledgments default to true to prevent data loss (#6999)__ [Indraja Babu](mailto:156961807+indrajababu@users.noreply.github.com) - Wed, 15 Jul 2026 09:34:27 -0500 Change the SQS source plugin default for acknowledgments from false to true. With the previous default, messages were deleted from SQS after being written to the internal buffer but before downstream delivery was confirmed. If the pipeline failed after buffering (sink unavailable, worker crash, processor error), messages were permanently lost with no recovery path. With acknowledgments enabled by default, messages are only deleted from SQS after the entire pipeline confirms successful end-to-end delivery. Changes: - Change default from false to true in SqsSourceConfig - Add WARNING log when acknowledgments is explicitly disabled - Add javadoc explaining data loss risk and visibility timeout implications - Add sqsMessagesDeletedWithoutAcknowledgment metric counter - Add integration tests verifying behavior with both modes under processing failure This is a BREAKING CHANGE. Users who explicitly require the previous best-effort behavior should set acknowledgments: false in their pipeline configuration. Users should ensure their SQS visibility timeout exceeds the expected end-to-end processing time to avoid duplicate delivery. Fixes #6998 Signed-off-by: Indraja Babu <indbabu@amazon.com> * __Add metadata for document routing to OpenSearch source (#6994)__ [fifthist](mailto:Overseer.O5.X@gmail.com) - Tue, 14 Jul 2026 07:34:07 -0700 Exposes the document routing value from search hits via the opensearch-document_routing event metadata attribute, mirroring the existing document id, index, and version metadata. Routing is only added when present so documents without custom routing are unaffected. Signed-off-by: Sergey Kojoian <skojoian.am@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * __Accept empty string as null in PluginModel deserializer and add Coerc… (#6915)__ [Siqi Ding](mailto:dingdd@amazon.com) - Mon, 13 Jul 2026 15:49:27 -0700 Accept empty string as null in PluginModel deserializer and add CoercionConfig for settings conversion Remove global CoercionConfig, keep targeted empty-string handling Remove the SERIALIZER_OBJECT_MAPPER CoercionConfig that globally coerced empty strings to null across all fields. The targeted VALUE_STRING check in the deserializer already handles bare YAML keys (e.g. stdout:) by treating empty string as null only at the plugin name level. Signed-off-by: Siqi Ding <dingdd@amazon.com> * __Fix thread safety in dissect processor by eliminating mutable shared state (#6858)__ [Shikhar Dhawale](mailto:51256166+shik-25@users.noreply.github.com) - Wed, 8 Jul 2026 09:39:52 -0700 * Fix thread safety in dissect processor by eliminating mutable shared state Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * Address review comments: make Field.key final, package-protect set Next, pre-size HashMaps, pre-sort AppendField lists, add ERROR logs, expand test coverage Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * Restore inline comments in DissectProcessorTest Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * Remove zip64 benchmarking artifact from build.gradle Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * Add debug logging when indirect field references are dropped in dissect processor Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * Address review comments: make template maps final, fix HashMap initial capacity in dissectText Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> --------- Signed-off-by: Shikhar Dhawale <shikhar_dhawale@apple.com> Co-authored-by: Shikhar Dhawale <shikhar_dhawale@apple.com> * __Add debug logging for every document id success (#6987)__ [Taylor Gray](mailto:tylgry@amazon.com) - Tue, 7 Jul 2026 18:05:51 -0500 Signed-off-by: Taylor Gray <tylgry@amazon.com> * __Add /ready endpoint to DataPrepperServer that returns 200 only after all pipeline sources have started. This prevents ALB from routing traffic to targets before the data ingestion port is listening, eliminating 502 errors during scale-up events. (#6977)__ [Siqi Ding](mailto:dingdd@amazon.com) - Tue, 7 Jul 2026 11:25:19 -0500 Signed-off-by: Siqi Ding <dingdd@amazon.com> * __Update the release process documentation for latest changes. (#6986)__ [David Venable](mailto:dlv@amazon.com) - Mon, 6 Jul 2026 12:28:35 -0700 Update the release process documentation for latest changes. We have made a number of significant changes to the release process. Some of these come from changes in the OpenSearch project and some are from additional automation we have added. This includes updates for how we coordinate with other parts of the OpenSearch project, steps that require the OpenSearch project to create branches and tags, and updates based on automation. I've also added additional information about releasing blog posts and updating the community. Signed-off-by: David Venable <dlv@amazon.com> * __Add Data Prepper 2.16.0 changelog. (#6973)__ [opensearch-ci](mailto:83309141+opensearch-ci-bot@users.noreply.github.com) - Mon, 6 Jul 2026 07:43:15 -0700 Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: dlvenable <293424+dlvenable@users.noreply.github.com> * __Publish a pre-release GitHub Release and use that for the Jenkins job. (#6974)__ [David Venable](mailto:dlv@amazon.com) - Thu, 2 Jul 2026 11:55:59 -0700 Publish a pre-release GitHub Release and use that for the Jenkins job. Per changes in the OpenSearch project, we can no longer modify repos using bots. So the new process is to publish a pre-release Release. Then the Jenkins job will use this pre-release. It's final step is to create a GitHub issue instead of publishing the Release. Use ncipollo/release-action instead of softprops/action-gh-release in the release action to keep consistency with other OpenSearch projects. Provide failure notifications for failing Jenkins jobs. Signed-off-by: David Venable <dlv@amazon.com> * __Updates the Claude command to generate release notes with instructions to look at the release blog. (#6952)__ [David Venable](mailto:dlv@amazon.com) - Wed, 1 Jul 2026 10:57:08 -0700 Using the release blog is a good way to be sure that the release notes have all the features we expect. It also is a good indication of what we want to highlight so we can have these promoted better in the release notes. Signed-off-by: David Venable <dlv@amazon.com> * __Adds release notes for Data Prepper 2.16.0. (#6951)__ [David Venable](mailto:dlv@amazon.com) - Wed, 1 Jul 2026 09:05:38 -0700 Adds release notes for Data Prepper 2.16.0. Signed-off-by: David Venable <dlv@amazon.com> * __Update fast-uri and brace-expansion to fix CVEs CVE-2026-6322,CVE-2026-6321 (#6969)__ [Krishna Kondaka](mailto:krishkdk@amazon.com) - Wed, 1 Jul 2026 07:31:18 -0700 Signed-off-by: Kondaka <krishkdk@amazon.com> * __Adds a new Claude command for generating the Thank You text and notes this command in the documentation for it. (#6950)__ [David Venable](mailto:dlv@amazon.com) - Wed, 1 Jul 2026 07:27:41 -0700 Signed-off-by: David Venable <dlv@amazon.com> * __Updates the next version to 2.17. (#6949)__ [David Venable](mailto:dlv@amazon.com) - Tue, 30 Jun 2026 15:53:14 -0700 Signed-off-by: David Venable <dlv@amazon.com> * __Onboard new backport-pr re-usable github workflow (data-prepper) (#6939)__ [Peter Zhu](mailto:zhujiaxi@amazon.com) - Tue, 30 Jun 2026 15:39:50 -0700 - Replace old backport workflow with reusable workflow from opensearch-build - Remove obsolete backport-related workflows Signed-off-by: Peter Zhu <zhujiaxi@amazon.com> * __Add classified error counters to CloudWatch Logs sink (#6943)__ [kiran536](mailto:veeravelly.kirankumar@gmail.com) - Tue, 30 Jun 2026 12:57:53 -0700 * Add classified error counters to CloudWatch Logs sink Introduces three per-class error counters (cloudWatchLogsAccessDenied, cloudWatchLogsResourceNotFound, cloudWatchLogsThrottled) so operators can alarm on specific failure modes rather than relying on the single opaque cloudWatchLogsRequestsFailed total. Signed-off-by: Kiran Kumar Veeravelly <kiran536@gmail.com> Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * Add forkEvery to integrationTest task for JVM isolation Prevents Mockito state from CloudWatchLogsIT (which requires real AWS credentials) from leaking into CloudWatchLogsMetricsIT when both test classes run in the same Gradle task. Signed-off-by: Kiran Kumar Veeravelly <kiran536@gmail.com> Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * Address review: robust AccessDenied matching and surface creation failures - Use contains("AccessDenied") instead of equals("AccessDeniedException") to handle both AWS error code variants (AccessDenied, AccessDeniedException) - Classify and count access-denied exceptions in createResources() so that permission problems during log group/stream creation are no longer invisible - Add isAccessDenied() and classifyAndCountResourceCreationFailure() helpers - Add unit tests for AccessDenied (no suffix), createLogStream access denied, createLogGroup access denied, and negative misclassification assertion Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * Gate resourceNotFound counter on user-owned provisioning Only increment cloudWatchLogsResourceNotFound when createLogGroup and createLogStream are both disabled. When either flag is on, the user delegated resource creation to the sink — a subsequent RNF is internal control flow, not an actionable user signal. Adds tests for both partial-flag permutations (createLogStream-only, createLogGroup-only) verifying the counter is suppressed. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> * Document new classified error counters in README Add cloudWatchLogsThrottled, cloudWatchLogsAccessDenied, and cloudWatchLogsResourceNotFound to the Metrics section with descriptions of when each counter is incremented and its semantic meaning. Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> --------- Signed-off-by: Kiran Kumar Veeravelly <kiran536@gmail.com> Signed-off-by: Kiran Kumar Veeravelly <veeravkk@amazon.com> Co-authored-by: Kiran Kumar Veeravelly <veeravkk@amazon.com>