The OpenTelemetry Collector Contrib contains everything in the opentelemetry-collector release, be sure to check the release notes there as well.

End User Changelog

🛑 Breaking changes 🛑

• exporter/clickhouse: Remove deprecated clickhouse.json feature gate. Users should set json: true in the exporter config directly. (#47888)

• extension/aws_logs_encoding: Remove deprecated format values and the vpc_flow_log config field. (#42901) The deprecated format values vpc_flow_log, elb_access_log, s3_access_log, cloudtrail_log, waf_log, and cloudwatch_logs_subscription_filter are no longer supported. Use vpcflow, elbaccess, s3access, cloudtrail, waf, and cloudwatch respectively. The deprecated vpc_flow_log config field has been removed; use vpcflow instead.

• pkg/ottl: Return errors when OTTL datapoint context setters are used on an incompatible data point type (#48384) For example, set(explicit_bounds, [1.0]) against a NumberDataPoint now returns an error rather than silently no-opping. Statements that were previously failing silently due to data point type mismatches will now surface as errors.

  Affected paths and the data point types they support:

  • value_double, value_int: NumberDataPoint
  • explicit_bounds, bucket_counts: HistogramDataPoint
  • scale, zero_count, positive, positive.offset, positive.bucket_counts, negative, negative.offset, negative.bucket_counts: ExponentialHistogramDataPoint
  • quantile_values: SummaryDataPoint
  • exemplars: NumberDataPoint, HistogramDataPoint, ExponentialHistogramDataPoint
  • count, sum: HistogramDataPoint, ExponentialHistogramDataPoint, SummaryDataPoint

• processor/filter: Promote processor.filter.defaultErrorModeIgnore feature gate to beta. (#47232) The default error_mode of the filter processor is now ignore instead of propagate. To restore the previous behavior, disable the feature gate with --feature-gates=-processor.filter.defaultErrorModeIgnore.

• processor/transform: Move the processor.transform.defaultErrorModeIgnore feature gate to beta. The default top-level error_mode is now ignore instead of propagate. (#48415) To revert to the previous default, disable the gate with --feature-gates=-processor.transform.defaultErrorModeIgnore.

• receiver/http_check: Fix timing metrics (httpcheck.dns.lookup.duration, httpcheck.client.connection.duration, httpcheck.tls.handshake.duration, httpcheck.client.request.duration, httpcheck.response.duration) always reporting 0 on fast networks where phase durations are sub-millisecond. Metrics now report values in nanoseconds instead of milliseconds. (#47257) Integer division truncated sub-millisecond durations to 0. Metrics now use nanoseconds as the unit, so a 500µs duration is reported as 500,000 rather than 0.

• receiver/jaeger: Remove stable gate receiver.jaeger.DisableRemoteSampling (#48616)

• receiver/pprof: Restructure config under remote, file, self, and server sections, and add HTTP push ingestion at POST /v1/pprof. (#48276) The top-level fields endpoint, include, collection_interval, initial_delay, block_profile_fraction, and mutex_profile_fraction are now nested under the corresponding remote, file, or self subsection. Multiple modes can be enabled on a single receiver instance, each with its own collection interval. A new server section enables push-mode ingestion of pprof data over HTTP.

🚩 Deprecations 🚩

• connector/metrics_as_logs: Rename component type from metricsaslogs to metrics_as_logs to follow snake_case naming convention. The old name is kept as a deprecated alias. (#48016)

• exporter/load_balancing: Rename the loadbalancing exporter to load_balancing. The old loadbalancing type remains available as a deprecated alias. (#45339)

• pkg/kafka/configkafka: Deprecate Kafka client config fields that became no-ops after the migration to franz-go. (#48260)

  • resolve_canonical_bootstrap_servers_only: franz-go has no direct equivalent to the associated Sarama config.
  • auth.sasl.version: franz-go negotiates the SASL handshake version automatically.

  Both fields are still accepted in configuration for backwards compatibility, but have no effect at runtime. They will be removed in a future release.

• processor/k8s_attributes: Deprecate deployment_name_from_replicaset and default deployment name extraction to the ReplicaSet name heuristic. (#48447)

• processor/resource_detection: Rename to resource_detection with deprecated alias resourcedetection (#48525)

🚀 New components 🚀

• processor/gen_ai_normalizer: Add a processor that normalizes GenAI telemetry attributes from OpenInference and OpenLLMetry to the official OTel GenAI Semantic Conventions. (#46069)

💡 Enhancements 💡

• cmd/opampsupervisor: Adds support for authenticator extensions in the supervisor when connecting to the upstream OpAMP server. (#47690) The Supervisor now supports the Bearer Token Authenticator Extension, Basic Auth Authenticator Extension, and OAuth2 Client Credentials Authenticator Extension. See the readme for more details.

• cmd/opampsupervisor: Adds a framework for the supervisor to utilize extensions from the collector ecosystem (#47690)

• cmd/opampsupervisor: Initial PR implementing package upgrades. Contains skeleton code that is unused. (#47298)

• connector/count: Support OTTL path context names in conditions. (#48316) Conditions for spans, span events, metrics, data points, logs, and profiles can now use context-prefixed paths (e.g. span.attributes["env"], resource.attributes["host"], metric.name). Existing un-prefixed paths continue to work; they are interpreted in the context of the enclosing block. It is recommend to update your configuration to the new syntax to avoid breaking changes in the future.

• connector/signal_to_metrics: Support OTTL context path prefixes in conditions and value expressions. (#48357) OTTL strings in conditions, value, count, and keys_expression fields may now use context path prefixes (e.g. span.attributes["foo"], datapoint.value_int, log.body, profile.duration_unix_nano). Unprefixed paths continue to work; it is highly recommend to use the new syntax to avoid breaking changes in the future.

• connector/span_metrics: Add an opt-in series_expiration setting to expire stale spanmetrics series without changing existing metrics_expiration behavior. (#44187)

• connector/sum: Support OTTL path-context names (e.g. span.attributes["foo"], resource.attributes["bar"], metric.name) in the conditions field. (#48329) Un-prefixed paths continue to work for now. If you are using un-prefixed paths, the updated statements will be printed on startup. It is highly recommended to switch to the new syntax to avoid breaking changes in the future.

• exporter/alertmanager: adds support for sending logs to Alertmanager. (#48099)

• exporter/elasticsearch: Implement experimental _noindex mapping hint to suppress indexing of individual signals. (#48310) This is an experimental feature and may change or be removed in future releases. When a span, span event, log record, or metric data point carries the _noindex value in its elasticsearch.mapping.hints attribute, the exporter skips emitting the bulk index operation for that document.

• exporter/google_cloud_storage: Add retry and sending queue configuration options (#48045) Introduces retry_on_failure and sending_queue support via exporterhelper. Also ensures that non-retryable GCS SDK errors correctly return permanent errors to halt the sending queue, preventing infinite retries. Includes documentation on the interaction between GCS SDK retries and exporter helper retries.

• exporter/honeycomb_marker: Support OTTL path-context names (e.g. log.body == "x", resource.attributes["service.name"] == "y") in markers[].rules.log_conditions. (#48325) Un-prefixed paths continue to work for now. If you are using un-prefixed paths, the updated statements will be printed on startup. It is highly recommended to switch to the new syntax to avoid breaking changes in the future.

• exporter/kafka: Add health reporting to kafka exporter. (#47293)

• exporter/kafka: Add per-signal message_key_from_metadata_key to derive the Kafka record key from client metadata. (#29433) Each signal (logs, metrics, traces, profiles) now accepts a message_key_from_metadata_key field that names a client metadata key whose value is used as the Kafka record key. This is mutually exclusive with the existing partition_* flags for the same signal. If the metadata key is absent or empty the record key is left nil.

• exporter/load_balancing: Add owner_account option to AWS Cloud Map resolver for cross-account namespace discovery (#47895) The owner_account field allows the AWS Cloud Map resolver to discover instances in namespaces shared from other AWS accounts using AWS RAM. This maps directly to the OwnerAccount parameter in the AWS Cloud Map DiscoverInstances API call.

• exporter/sumologic: Modify the sumologic exporter's default batching configuration (#47820)

• extension/aws_logs_encoding: Add cloudwatch.streams for the cloudwatch format to route subscription-filter events to different inner encoding extensions based on logGroup/logStream patterns. (#46458) Each stream accepts a name, an encoding (component ID of the inner extension), optional log_group_pattern / log_stream_pattern matchers, and an optional payload mode (message or envelope). Known names (vpcflow, cloudtrail, lambda, waf, rds, eks, apigateway) carry default patterns and payload modes reflecting AWS conventions. Routing is per envelope, so CloudWatch payloads aggregating multiple log groups (e.g. from Firehose) are dispatched correctly. Envelopes that match no stream fall back to the default subscription-filter behavior.

• pkg/stanza: Respect the "SynchronousLogEmitter" feature gate in container parser operator. (#47828) The container parser operator now respects the stanza.synchronousLogEmitter feature gate, using the synchronous log emitter when enabled to prevent possible data loss. Enabling the feature gate disables batching within the container parser.

• processor/cardinality_guardian: Add core processor logic, HLL sharding, enforcement modes, and inline spatial reaggregation to cardinalityguardianprocessor. (#47368)

• processor/drain: Add snapshot persistence via storage extension to save and restore the Drain tree across restarts. (#48075) New config fields: storage (ID of a storage extension) and save_interval (periodic snapshot interval). With shared storage, periodic saves allow new instances in a scaled deployment to inherit a trained tree.

• processor/gen_ai_normalizer: Rename component type from genainormalizer to gen_ai_normalizer. (#46069)

• processor/gen_ai_normalizer: Add the openllmetry source for normalizing OpenLLMetry (Traceloop) GenAI attributes. (#46069)

• processor/gen_ai_normalizer: Implement attribute and value normalization for the OpenInference source. (#46069)

• processor/k8s_attributes: Improve k8s.cronjob.name extraction (#44831) Improve CronJob name extraction to use the Job informer when already running. Otherwise use heuristic to match the Job name's and check if the last 8-digits are a valid timestamp closely matching the pod creation time (24h window).

• processor/log_dedup: Support OTTL path-context names (e.g. log.attributes["foo"], resource.attributes["bar"]) in the conditions field. (#48266) Un-prefixed paths continue to work for now. If you are using un-prefixed paths, the updated statements will be printed on startup. It is highly recommended to switch to the new syntax to avoid breaking changes in the future.

• processor/schema: Promote schema processor from development to alpha stability and add to contrib distribution (#47428)

• processor/schema: Add migration mode to preserve original attributes alongside renamed ones during schema version changes (#47420)

• processor/schema: Add optional storage extension support to persist fetched schema files across collector restarts (#48222)

• processor/tail_sampling: Support OTTL path-context names (e.g. span.attributes["foo"], resource.attributes["bar"], spanevent.name) in the ottl_condition policy. (#48330) Un-prefixed paths continue to work for now. If you are using un-prefixed paths, the updated statements will be printed on startup. It is highly recommended to switch to the new syntax to avoid breaking changes in the future.

• processor/tail_sampling: Switch the rate_limiting policy to a token bucket and add burst_capacity to allow short bursts of traffic above the sustained spans-per-second rate. (#48226) If burst_capacity is not set, it defaults to 2x spans_per_second, mirroring the bytes_limiting policy.

  Existing configurations will see slightly different behavior: the bucket starts with a default burst of 2x spans_per_second tokens. A trace whose span count equals spans_per_second can now be sampled (the previous strict < comparison meant spans_per_second: 1 would never sample). Set burst_capacity explicitly to match the previous per-second cap if needed.

• provider/s3: Add support for Amazon S3-compatible services (e.g. MinIO) using path-style endpoints and insecure HTTP URIs. (#47595)

  • Hosts with non-Amazon DNS suffixes are now treated as path-style endpoints (s3://host/bucket/key) instead of virtual-hosted-style.

• receiver/awscloudwatch: Add metrics support to the AWS CloudWatch receiver. (#47330)

• receiver/file_log: Improves file-reading efficiency by evicting previously read data from the OS page cache. (#48273) Clears the cache on Linux; acts as a no-op on unsupported platforms.

• receiver/host_metrics: Warn and disable process.handles at startup on platforms where it is not supported, instead of logging an error every scrape cycle. (#47095) process.handles is only supported on Windows. Previously, enabling it on other platforms would spam "only supported on Windows" errors on every scrape cycle for every process. The factory now checks for this at construction, logs a single warning, and disables the metric so the hot path is never exercised.

• receiver/host_metrics: Warn and disable process.paging.faults at startup on platforms where it is not supported, instead of logging an error every scrape cycle. (#47095) process.paging.faults is only supported on Linux. Previously, enabling it on other platforms would spam "not implemented" errors on every scrape cycle for every process. The factory now checks for this at construction, logs a single warning, and disables the metric so the hot path is never exercised.

• receiver/host_metrics: Warn and disable process.context_switches at startup on platforms where it is not supported, instead of logging an error every scrape cycle. (#47095) process.context_switches is only supported on Linux. Previously, enabling it on other platforms would spam "not implemented" errors on every scrape cycle for every process. The factory now checks for this at construction, logs a single warning, and disables the metric so the hot path is never exercised.

• receiver/host_metrics: Warn and disable process.signals_pending at startup on platforms where it is not supported, instead of logging an error every scrape cycle. (#47095) process.signals_pending is only supported on Linux (via /proc//limits). Previously, enabling it on other platforms would spam "not implemented" errors on every scrape cycle for every process. The factory now checks for this at construction, logs a single warning, and disables the metric so the hot path is never exercised.

• receiver/http_check: Enables dynamic metric reaggregation in the HTTP Check receiver. This does not break existing configuration files. (#46358)

• receiver/k8s_objects: Add top-level interval field as a fallback default pull interval for all pull-mode objects. (#48452) When interval is set at the top-level k8s_objects config, it applies to all pull-mode objects that do not set their own per-resource interval. The resolution order is: per-resource interval → top-level interval → built-in default (1h). Watch-mode objects are unaffected. Existing configurations continue to work without change.

• receiver/kubelet_stats: Add optional k8s.container.ephemeral_storage.usage metric to kubeletstats receiver. (#47601)

• receiver/kubelet_stats: Enables dynamic metric reaggregation capability (#46363)

• receiver/ntp: Enable scraper reaggregation for the NTP receiver (#46370)

• receiver/oracledb: Add new attributes to db.server.query_sample event for enhanced session monitoring (#48216) Add 4 new attributes to the db.server.query_sample event: oracledb.query.wait_time, oracledb.query.started, oracledb.session.started, and oracledb.session.duration. Timestamps are in UTC ISO 8601 format. These attributes enable better monitoring of query wait times, execution timing, and session lifecycle.

• receiver/oracledb: Adds blocking session and lock attributes to db.server.query_sample events for Oracle blocking session detection. Blocking attributes are always emitted (empty string / 0 when not blocked). (#48143) New attributes under oracledb.blocking.*: oracledb.blocking.blocker.sid, oracledb.blocking.blocker.root_sid, oracledb.blocking.blocker.state, oracledb.blocking.start_time, oracledb.blocking.wait_duration, oracledb.blocking.lock.type, oracledb.blocking.lock.mode, oracledb.blocking.object.owner, oracledb.blocking.object.name. oracledb.blocking.start_time provides an estimated UTC timestamp (RFC3339) of when blocking began. Filters to S.TYPE = 'USER' to exclude SYS/background sessions. Idle blockers (INACTIVE sessions holding locks) are captured. All blocking attributes are always present on every db.server.query_sample event; empty string / 0 values indicate no active blocking.

• receiver/oracledb: Add operational metrics for data dictionary hit ratio, recycle bin size, and storage usage (#47303) Add 4 new opt-in metrics (disabled by default, stability: development): oracledb.data_dictionary.hit_ratio, oracledb.recycle_bin.limit, oracledb.storage.usage, and oracledb.storage.utilization.

• receiver/prometheus_remote_write: Performance improvements while processing remote write requests. (#47850)

• receiver/saphana: Enable the re-aggregation feature for the SAP HANA receiver. (#46378)

• receiver/sqlserver: Add sqlserver.query.last_started and db.namespace attributes to db.server.top_query events. (#48217) The db.server.top_query log records now include:

  • sqlserver.query.last_started: ISO 8601 timestamp of when the query last started executing.
  • db.namespace: the database name, populated via DB_NAME(st.dbid) in the underlying SQL query.

• receiver/vcenter: Add max_query_metrics option to batch QueryPerf requests, preventing vCenter resource exhaustion in large deployments. (#47941)

• receiver/windows_event_log: Add event_driven_scraping config option (#48235) The new event_driven_scraping config option enables the event-driven scraping model on the windows event log input (and the windowseventlog receiver) without requiring the stanza.windows.eventDrivenScraping feature gate. Event-driven scraping is enabled when either the config option or the feature gate is set. This configuration option should be treated as experimental until the feature gate has been promoted from alpha stability.

• receiver/windowsservice: Advancement of windows service receiver from development to alpha (#48176)

🧰 Bug fixes 🧰

• connector/service_graph: Fix client-side exponential histograms being emitted with a zero timestamp, causing backends like Grafana Cloud Mimir to reject them. (#48049) collectClientLatencyMetrics omitted SetTimestamp on the exponential-histogram data point, so the zero-value timestamp leaked through. The server-side function and the client-side explicit-bucket branch were already correct.

• exporter/datadog: Honor tls.insecure_skip_verify for logs flowing through datadogexporter (#48523)

• exporter/datadog: Fix percentile aggregations collapsing to 0 for OTLP explicit-bucket histograms whose first non-empty bucket is (0, B]. (#48478) Affects the default boundary set used by Micrometer's OTLP registry and many OTel SDKs ([0, 5, 10, 25, 50, 75, 100, 250, …]). With high-cardinality tagging and short delta intervals, per-bucket counts of 1 or 2 made the zero-bin leak dominate, collapsing p50…p99 aggregations to 0. avg, sum, min, max, and count were not affected.

• extension/sumologic: Make metadata update process async to prevent collector startup failure when metadata endpoint is unavailable (e.g. Windows service starting before network initialization). (#48450)

• pkg/translator/splunk: Propagate the OTLP log record EventName field into the Splunk HEC event under the configured otel_to_hec_fields/name key (default otel.log.name). (#48048)

• processor/k8s_attributes: Preserve an upstream-set container.image.tags resource attribute when processor.k8sattributes.EmitV1K8sConventions is enabled, matching the behavior of container.image.tag and container.image.repo_digests. (#45869)

• processor/k8s_attributes: Re-enable otelcol.k8s.pod.association metric with low-cardinality pod_identifier attribute. The attribute now encodes the association source type/name (e.g. "connection", "resource_attribute/k8s.pod.ip") rather than actual pod identifier values, preventing unbounded metric time series growth. (#47669)

• processor/schema: Enforces that all prefetched schemas are completed before starting. (#48258)

• receiver/azure_blob: Delete blobs only upon successful processing (#48211) Blobs that fail to be processed will now remain in the container and will be retried on the next poll. Users with consistently failing pipelines may see container growth or duplicate emissions on retry.

• receiver/azure_monitor: fix(azuremonitorreceiver): prevent duplicate sample timestamps in batch scraper (#47938)

• receiver/cloudflare: Preserve array-typed Cloudflare log fields as OTLP attributes instead of dropping them. (#46716)

• receiver/iis: The IIS receiver was failing to start when a disabled metric was not present in the system (#48609) The IIS receiver was attempting to create watchers for all metrics, even those that were disabled. This was causing the receiver to fail to start if any of the disabled metrics were not present in the system, which is a valid scenario for some of the IIS metrics. Now, the receiver will only attempt to create watchers for enabled metrics, so it should start successfully as long as all enabled metrics are present in the system.

• receiver/mysql: Exclude collector's own queries from Query Sample collection to prevent feedback loop. (#47904) Adds a /* otel-collector-ignore */ comment prefix to the Query Sample SQL template and filters it out via a NOT LIKE clause on statement.sql_text.

• receiver/oracledb: Exclude the collector's own monitoring queries from Query Sample collection using native session ID filtering. (#47366) Uses native session evaluation (AND S.SID != SYS_CONTEXT('USERENV', 'SID')) to filter the collector's active queries from V$SESSION. This bypasses Go connection pool limitations and avoids the performance overhead of text-matching on VARCHAR2 or CLOB fields. Top Query collection is no longer filtered, so collector-originated statements can still appear there for troubleshooting.

• receiver/podman_stats: Fix incorrect network I/O metric recording (#48313) The podman receiver was incorrectly recording network I/O metrics, with the received bytes and transmitted bytes swapped. This fix corrects the recording of these metrics to ensure accurate telemetry data is collected.

• receiver/prometheus_remote_write: Fix otelcol_receiver_accepted_metric_points reporting resource count instead of data point count. (#48306)

• receiver/prometheus_remote_write: Fix broken trace propagation to downstream pipeline components. (#48548)

• receiver/prometheus_remote_write: Properly handle NaN native histograms, and drop points associated with the native histogram overflow buckets. (#47728)

• receiver/sqlserver: Exclude collector's own queries from Query Sample collection to prevent feedback loop. (#47332) Adds a /* otel-collector */ comment prefix to the Query Sample SQL template and filters it out via a WHERE clause on sys.dm_exec_sql_text. Also excludes the current session via @@SPID. Top Query collection is intentionally left unfiltered so users can see the monitoring cost.

• receiver/syslog: Support days with leading zeros in RFC 3164 syslogs (#47665)

• receiver/yang_grpc: Fix setting metric attributes from context bag. (#48568)

• scraper/zookeeper: Fix zk_avg_latency metric being silently dropped on Zookeeper 3.7+ where the value is reported as a float. (#47320)

API Changelog

💡 Enhancements 💡

• extension/file_storage: Implement storage.Walker interface to allow iterating over all stored keys with deferred operations (#47755)

• internal/kafka: Remove the dependency on github.com/IBM/sarama from all Kafka components. (#48260) The Kafka exporter, receiver, metrics receiver, and topics observer have all been migrated to github.com/twmb/franz-go. The remaining sarama-based helpers in internal/kafka are removed, and protocol_version validation in pkg/kafka/configkafka now uses franz-go's kversion package.

• pkg/pdatatest: Add /exists operator support to pmetricassert (#48079)

• pkg/pdatatest: Introduce pmetricassert package for MTS-focused YAML metric assertions (#48079)

• receiver/file_log: Improves file-reading efficiency by evicting previously read data from the OS page cache. (#48273) Clears the cache on Linux; acts as a no-op on unsupported platforms.

• receiver/http_check: Enables dynamic metric reaggregation in the HTTP Check receiver. This does not break existing configuration files. (#46358)

We are thrilled to welcome our first-time contributors to this project. Thank you for your contributions @theGarimaAgrawal, @rdner, @jeffreylimnardy, @alien2003, @harshitt13, @thaJeztah, @iahsanGill, @sjparkinson, @shailygupta, @avadla, @sjyothi54, @yogeswaran, @VinozzZ ! 🎉