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

Unmaintained Components

These components are marked as Unmaintained and will eventually be removed from our community distributions. If you depend on these components we need your help to support them.

• Simple Prometheus receiver

End User Changelog

🛑 Breaking changes 🛑

• all: Remove the wavefront receiver after a period of deprecation (#51397)

• connector/signal_to_metrics: The default error mode has been changed from propagate to ignore to ensure valid OTLP payloads are not dropped (#48419)

• exporter/elasticsearch: Remove the deprecated flush and num_workers configuration settings. (#42718) Migrate num_workers to sending_queue::num_consumers, flush::interval to sending_queue::batch::flush_timeout, and flush::bytes to sending_queue::batch::max_size with sending_queue::batch::sizer set to bytes. This also removes Config.NumWorkers, Config.Flush, and FlushSettings from the exported Go API.

• exporter/elasticsearch: Ingest OTel profiles as Elasticsearch (9.6.0) datastreams (#50589) In otel mapping mode, profiling signals are now written to OTel-native datastreams (e.g. profiling-events-all.otel-default) that require Elasticsearch 9.6.0 or later. Sending profiles to an older cluster will result in index_not_found_exception errors and data loss. Users on Elasticsearch < 9.6.0 must restrict the exporter to ECS mode via mapping.allowed_modes: [ecs].

• exporter/elasticsearch: Move OTel profiling serializer to ecsserializer package and wire ecs mapping mode to use it for profiles. (#50589)

• pkg/ottl: Remove the stable ottl.PanicDuplicateName and ottl.contexts.enableOTelColContext feature gates. (#44630, #46437)

• pkg/prometheus: Promote the pkg.translator.prometheus.PermissiveLabelSanitization feature gate from alpha to beta, enabling it by default. (#50429) Labels starting with a single underscore (e.g. _foo) will no longer be prefixed with key_ by default. To restore the previous behavior, disable the feature gate: --feature-gates=-pkg.translator.prometheus.PermissiveLabelSanitization.

• processor/adaptive_tail_sampling: Add explicit root-scoped fingerprint selectors (root.resource., root.scope., root.span., root.any.) and require root. to name an origin. (#51304) root. is now a position prefix that composes with the resource, scope, span and any origins, so a fingerprint can read the root span's resource-level attributes (root.resource.attributes["service.name"]) or their union (root.any.attributes[...]). Bare root.attributes["<name>"] no longer parses; it previously read only the root span's span attributes, so resource attributes such as service.name came back <missing>. Migrate root.attributes[...] to root.span.attributes[...] for the same behaviour, or to root.resource. / root.any. to reach the attributes that were unreachable before.

• processor/cardinality_guardian: move units to singular form (#45270)

• processor/signing: Wrap attribute scalar values in type-tagged objects ({"intValue":"…"}, {"stringValue":"…"}, etc.) to eliminate scalar type collisions in the JCS-canonical signed payload. (#50884) Previously, PutInt("k", 123) and PutStr("k", "123") produced identical canonical JSON, so one signature covered both values. Each scalar type is now wrapped in a single-key object keyed by type name (matching OTLP/JSON conventions), making every type distinct. Existing signatures are invalidated by this change. SeverityNumber and SeverityText are no longer included in the signed payload; the audit logging spec marks them as SHOULD NOT.

• processor/signing: Adds OpenBAO and Kubernetes Secrets as newKeyMaterialProvider options to the signing processor. (#50079, #51055) Additionally, supports a Kubernetes Secret, or an OpenBao/Vault KV secret. Sets audit.integrity.signer = "collector" to indicate the Tier-2 Collector produced the integrity proof.

• processor/transform: Remove the stable processor.transform.defaultErrorModeIgnore feature gate. (#51242)

• receiver/kafka: Change the default for metadata::retry::max from 3 to 20 to match franz-go's default. (#50116) The field was unwired since the migration to franz-go, so every collector has effectively been running with franz-go's default of 20. The new default preserves that behavior now that the field is applied. Only configurations that set metadata::retry::max explicitly change behavior: the configured value is now honored.

• receiver/kafka_metrics: Change the default for metadata::retry::max from 3 to 20 to match franz-go's default. (#50116) The field was unwired since the migration to franz-go, so every collector has effectively been running with franz-go's default of 20. The new default preserves that behavior now that the field is applied. Only configurations that set metadata::retry::max explicitly change behavior: the configured value is now honored.

• receiver/prometheus: Promote the receiver.prometheusreceiver.IgnoreScopeInfoMetric feature gate to stable (#47312) The otel_scope_info metric is no longer used for scope attribute extraction and is converted like any other metric. Scope attributes are populated from otel_scope_-prefixed metric labels instead. The feature gate can no longer be disabled and will be removed in v0.164.0.

🚩 Deprecations 🚩

• connector/grafana_cloud: Rename component type from "grafanacloud" to "grafana_cloud" to follow snake_case naming convention. The old name is kept as a deprecated alias. (#48017)

• receiver/aws_s3: Rename awss3 receiver to aws_s3 with deprecated alias awss3 (#45339)

• receiver/kafka: Deprecate metadata::full, which became a no-op after the migration to franz-go. (#50116) franz-go has no full-metadata mode; it only fetches metadata for the topics the receiver consumes from. The field is still accepted in configuration for backwards compatibility, but has no effect at runtime. It will be removed in a future release.

• receiver/kafka_metrics: Deprecate metadata::full, which became a no-op after the migration to franz-go. (#50116) franz-go has no full-metadata mode; it only fetches metadata for the topics the scrapers query. The field is still accepted in configuration for backwards compatibility, but has no effect at runtime. It will be removed in a future release.

• receiver/kubelet_stats: Deprecate k8s.pod.cpu.node.utilization, k8s.pod.memory.node.utilization, k8s.container.cpu.node.utilization, and k8s.container.memory.node.utilization metrics. (#50857) These metrics are deprecated following the decision in the K8s SemConv SIG (https://github.com/open-telemetry/semantic-conventions/issues/2768). They will be removed in a future release.

🚀 New components 🚀

• exporter/nats: Add the NATS exporter to export traces, metrics, and logs to a NATS server (skeleton; in development). (#39540)
• extension/file_telemetry_policy: Add filetelemetrypolicy extension skeleton. (#50965)
• extension/sd_notify: Promote sd_notify extension stability from development to alpha. (#49607)
• processor/telemetry_policy: Add the initial skeleton for the telemetry_policy processor. (#50965)

💡 Enhancements 💡

• cmd/opampsupervisor: Enable keepalive by default (#51159)

• connector/exceptions: Add logs_to_logs and logs_to_metrics support, and fix the traces_to_logs edge to emit spec-conformant exception LogEvents (event.name, required exception.* attributes, and full resource attribute propagation). (#50900)

• connector/failover: Introduces condition.error.contains which allows the connector to failover based on error type. (#44518)

• exporter/azure_monitor: Enable keepalive by default (#51178)

• exporter/coralogix: Enable keepalive by default (#51181)

• exporter/datadog: Emit otel.datadog_exporter.metrics.running.azurecontainerapps tagged with replica, name, subscription_id, and resource_group for Azure Container Apps workloads sending metrics via the Datadog exporter. (#49615)

• exporter/datadog: Emit running metrics for Azure App Services, Azure Functions, Cloud Run services, and Cloud Run Functions workloads sending metrics through the Datadog exporter. (#51501)

• exporter/honeycomb_marker: Enable keepalive by default (#51180)

• exporter/load_balancing: Add randomness routing key that routes spans by the OTel tracestate randomness value (ot=rv), falling back to trace ID randomness when absent, so traces sharing an explicit randomness value are routed to the same backend. (#49660)

• exporter/logicmonitor: Enable keepalive by default (#49316)

• exporter/sematext: Enable keepalive by default (#49316)

• exporter/syslog: Add support for exporting to Unix datagram sockets such as /dev/log. (#50867)

• extension/awsproxy: Enable keepalive by default (#49316)

• extension/datadog: Enable keepalive by default (#49316)

• extension/health_check: Enable keepalive by default (#51175)

• extension/healthcheckv2: Enable keepalive by default (#51175)

• extension/http_forwarder: Enable keepalive by default (#51182)

• extension/jaegerremotesampling: Enable keepalive by default (#49316)

• extension/json_log_encoding: Add the opt-in parse_ints setting to preserve JSON integer literals within the int64 range. (#51299) Applies to array and single-document/NDJSON input, including nested objects and arrays. Decimal, exponent, and out-of-range integer literals are still parsed as float64. The option defaults to false and adds processing overhead when enabled.

• extension/mcp: Enable keepalive by default (#51177)

• extension/pebble_tail_storage: Add extension telemetry counters for Pebble tail storage operations and read-path failures (#49697) Adds the following metrics:

  • otelcol_extension_pebble_tail_storage_operations (operation=append|take|delete, outcome=success|failure)
  • otelcol_extension_pebble_tail_storage_read_errors

• extension/remotetap: Enable keepalive by default (#49316)

• extension/sumologic: Enable keepalive by default (#49316)

• pkg/ottl: Promote the ottl.set.allowNil feature gate to stable. (#49741)

• processor/adaptive_tail_sampling: Add sampler_request_count, sampler_keyspace_size, and sampler_burst_count metrics exposing the internal performance of adaptive_percentage and adaptive_throughput samplers, labelled by rule, sampler_type, and sampler_algorithm. (#49311) Metrics are labelled by rule, sampler_type (adaptive_percentage, adaptive_throughput), and sampler_algorithm (ema, windowed), matching the rule's sampler config, so keyspace growth, request bursts, and convergence state can be compared per rule. sampler_burst_count is not emitted for adaptive_throughput rules using the windowed algorithm, which doesn't track this counter.

• processor/adaptive_tail_sampling: Default max_keys to 500 when omitted; explicitly setting it to 0 still means unlimited. (#49311)

• processor/interval: Add delta temporality metrics support. (#50845)

• processor/lookup: Promote lookup processor to alpha stage (#51199)

• processor/remotetap: Enable keepalive by default (#49316)

• processor/resource_detection: Add Azure Functions resource detector (#50633)

• processor/rolling_span_latency: Implement EWMA-based rolling latency baseline tracking and slow/very_slow span labeling. (#50260)

• processor/transform: Add extract_avg_metric OTTL function to transform processor (#49238)

• receiver/aws_s3: Support coarser time-steps in partition formats (#51038) Adds support for day-partitioned time slices when reading partition formats without minute/hour granularity.

• receiver/awscontainerinsightreceiver: Enable keepalive by default (#49316)

• receiver/awsecscontainermetrics: Enable keepalive by default (#49316)

• receiver/awsfirehose: Enable keepalive by default (#49316)

• receiver/awsxray: Enable keepalive by default (#49316)

• receiver/azure_functions: Enable keepalive by default (#49316)

• receiver/azure_monitor: Add support for filtering Azure resources by tags. (#50959) Filter resources by tag name and optionally by tag value.

• receiver/azure_monitor: Emit the timegrain datapoint attribute on the ARM collection path, matching the batch API path (#50847) Datapoints collected through the batch API carried a timegrain attribute while the same metrics collected through the ARM API did not, so a consumer could not tell a PT1M series from a PT1H one on that path. Both scrapers now attach the attribute from the same composite key.

• receiver/cloudflare: Enable keepalive by default (#49316)

• receiver/collectd: Enable keepalive by default (#49316)

• receiver/datadog: Enable keepalive by default (#49316)

• receiver/file_stats: move all attributes and metrics to beta (#51260)

• receiver/github: Enable keepalive by default (#49316)

• receiver/gitlab: Enable keepalive by default (#49316)

• receiver/gitlab: Add gitlab.project.id resource attribute (#48993)

• receiver/google_cloud_spanner: Add missing label operations_by_table_json_string to top minute transaction stats in googlecloudspannerreceiver. (#51057) Syncs the OTel receiver with SPANNER_SYS tables to bring the parity between GCP transaction insights page and OTel receiver.

• receiver/google_cloud_spanner: Add query_type label to top minute query stats in googlecloudspannerreceiver. (#50843) Syncs the OTel receiver with SPANNER_SYS tables to bring the parity between GCP Query insights page and OTel receiver.

• receiver/haproxy: Enable keepalive by default (#49316)

• receiver/haproxy: Move all metrics and attributes to beta (#51261) Also, change all units to singular form per #45270

• receiver/http_check: Enable keepalive by default (#49316)

• receiver/huaweicloudcesreceiver: Enable keepalive by default (#49316)

• receiver/influxdb: Enable keepalive by default (#49316)

• receiver/jaeger: Enable keepalive by default (#49316)

• receiver/journald: Add "convert_to_semantic_conventions" config option to map well-known journald fields to OpenTelemetry semantic convention attributes (#7298) See the Semantic Conventions Mapping section of the receiver README for details on what convert_to_semantic_conventions does.

• receiver/kafka: Link the receive span to the trace context in Kafka record headers, using the propagators configured in service::telemetry::traces::propagators. (#51077)

• receiver/libhoney: Enable keepalive by default (#49316)

• receiver/loki: Enable keepalive by default (#51183)

• receiver/mongodb: Add replica set topology and oplog metrics (#50654) Seven opt-in metrics, disabled by default, emitted only by replica set members.

• receiver/mysql: Add server.address and server.port resource attributes (#50967) Both attributes are emitted by default. When the receiver connects over loopback (for example localhost or 127.0.0.1), server.address reports the host name of the machine running the collector, matching how service.instance.id already resolves its host. mysql.instance.endpoint and existing service.instance.id values are unchanged. Disable the attributes via resource_attributes.server.address.enabled: false and resource_attributes.server.port.enabled: false.

• receiver/mysql: Add a new db.server.query_plan event that reports the query execution plan on a record of its own, so it can be filtered, routed or dropped independently of the query statistics. (#51281) Disabled by default, and requires db.server.top_query or db.server.query_sample. Enabling it removes mysql.query_plan from those two events, which keep mysql.query_plan.hash as the join key. db.server.top_query also reports db.namespace, whether or not the new event is enabled.

• receiver/nsxt: Enable keepalive by default (#49316)

• receiver/ntp: Move ntp receiver metric to beta (#51380)

• receiver/oracledb: Add a new db.server.query_plan event that reports the query execution plan on a record of its own, so it can be filtered, routed or dropped independently of the query statistics. (#51065) Disabled by default, and requires db.server.top_query.

• receiver/oracledb: Add the oracle.db.edition resource attribute. (#51292) Disabled by default. opt in by setting resource_attributes.oracle.db.edition.enabled: true in the receiver config.

• receiver/oracledb: Collect system and resource limit metrics when the receiver is connected directly to a PDB, such as on AWS RDS Oracle (#50147) Previously, system and resource limit metrics were not reported when connecting straight to a PDB.

• receiver/oracledb: Add the db.system.name attribute to the db.server.query_plan and db.server.session.wait_sample events. (#51065)

• receiver/postgresql: Add postgresql.backend.connection.start attribute to db.server.query_sample (#50769)

• receiver/postgresql: Add db.system.version resource attribute. (#51288) Disabled by default. opt in by setting resource_attributes.db.system.version.enabled: true in the receiver config.

• receiver/postgresql: Emit the server.address and server.port resource attributes regardless of the receiver.postgresql.useOTelSemconv feature gate, and report a loopback endpoint as the collector's host name in server.address. (#50889) Both attributes were previously emitted only with the gate enabled and are additive, so the resource model the gate selects is otherwise unchanged. Loopback resolution matches how service.instance.id already resolves the same endpoint; non-loopback endpoints are unaffected, as is transport: unix, which continues to report the socket path.

• receiver/postgresql: Add a new db.server.query_plan event to report the query execution plan separately. (#51301) Disabled by default, and requires db.server.top_query. When enabled, db.server.top_query stops carrying postgresql.query_plan.

• receiver/prometheus: Enable keepalive by default (#49316)

• receiver/prometheus: Use value types and pre-allocate slices when processing scraped samples and building OTLP metrics. (#51133)

• receiver/prometheus_remote_write: Enable keepalive by default (#49316)

• receiver/rabbitmq: Add the disabled-by-default rabbitmq.cluster.name resource attribute. (#50168)

• receiver/receiver_creator: Add a joinHostPort function for receiver template expressions. (#50969)

• receiver/signalfx: Enable keepalive by default (#49316)

• receiver/skywalking: Enable keepalive by default (#49316)

• receiver/splunk_hec: Enable keepalive by default (#49316)

• receiver/sqlserver: Add db.system.version resource attribute. (#51194) Only available in direct connection mode. Disabled by default.

• receiver/sqlserver: Add db.namespace and db.system.name to the db.server.query_plan event, so all four log events carry them. (#50629)

• receiver/tcp_check: move metric to alpha stability (#51381)

• receiver/vcenter: Add proxy_url to configure a per-receiver proxy for vSphere SDK connections. (#51187) The http, https, socks5 and socks5h schemes are supported. When proxy_url is unset the process-wide HTTP_PROXY/HTTPS_PROXY/NO_PROXY environment variables continue to apply, so existing behavior is unchanged. Because the setting is scoped to a single receiver instance, several vcenter/<name> receivers in one collector can now each reach their vCenter through a different proxy.

• receiver/webhook_event: Enable keepalive by default (#49316)

• receiver/zipkin: Enable keepalive connections by default (#51139)

🧰 Bug fixes 🧰

• all: Exclude datadogconnector, datadogexporter, and datadogextension on solaris (#51329)

• cmd/opampsupervisor: Preserve the health last reported to the OpAMP server when an accepted connection settings offer replaces the OpAMP client, instead of reporting a running agent as unhealthy (#51009) The supervisor.agent.health_status metric now also tracks the health reported when no config is present and at shutdown, which previously bypassed it.

• cmd/opampsupervisor: Open agent.log in append mode for external log rotation (Linux + Windows) (#50376) This makes external copytruncate-style log rotation (e.g. logrotate) work correctly instead of the file's size reverting on the next write, on both non-Windows (O_APPEND) and Windows (a FILE_APPEND_DATA handle, since Go's O_APPEND does not survive handle inheritance to the managed agent process on Windows).

• cmd/opampsupervisor: Fix stopping a Collector that had just started on Windows sometimes taking the full stop grace period and ending in a forced kill. (#51499)

• cmd/telemetrygen: Prevent unbounded attribute growth and declining metric generation rates when using --unique-timeseries. (#49950)

• connector/grafana_cloud: Make k8s.node.name the first entry in the default host_identifiers list so that hosts reported by several telemetry sources are counted once (#51502) The default is now ["k8s.node.name", "host.id", "k8s.node.uid"].

• exporter/doris: Fix materialized view and scheduled job creation on Doris 4.x, and add an is_root column plus a <traces>_summary materialized view to the traces table. (#51265) Existing tables are not altered; the new column and view only apply to tables created with create_schema: true.

• exporter/elasticsearch: Fix downsampled profiling events (profiling-events-5powNN) being silently dropped instead of sent to Elasticsearch. (#51222)

• exporter/kafka: Apply metadata::retry::backoff, which was silently ignored since the migration to franz-go. (#50116) metadata::retry::backoff is the minimum wait between retries; each successive retry doubles the wait, with jitter applied, capped at max(5s, backoff). On the exporter this paces produce retries. Produce retry counts remain governed by retry_on_failure and timeout.

• exporter/otelarrow: Recover from transient stream connection errors instead of permanently losing stream capacity. (#50254) Any failure to establish a stream was previously treated as evidence that the endpoint does not support OTel-Arrow, however some transient errors like a timeout could also surface on these code paths and violate that assumption. This resulted in an unhandled partial failure case where the worker states associated with those failed streams would become orphaned with no routine processing their inputs. They would also appear to be unloaded, attract requests, and then permanently wedge the sending goroutine because nothing writes back to it.

• exporter/prometheus: Remove every series ended by a staleness marker when several arrive in the same metric. (#51289)

• exporter/pulsar: Exclude pulsarexporter on solaris (#51330)

• extension/google_cloud_logentry_encoding: Fix decoding of log messages larger than 64KiB, which previously failed with "bufio.Scanner: token too long". (#51014)

• extension/k8s_observer: Format IPv6 Kubernetes port endpoints correctly. (#50969)

• extension/pebble_tail_storage: Exclude pebbletailstorageextension on solaris (#51327)

• pkg/kafka/configkafka: Fix missing AWS MSK SASL validation and incomplete OAUTHBEARER error message (#50934) Validate that the region is not empty when the AWS MSK mechanism is used, and add the missing OAuth mechanism to the default switch case error message.

• pkg/ottl: Reject lambda expressions at parse time when they are not passed to a function argument that accepts them. (#51560)

• pkg/stanza: Call the Windows Event Log API through *windows.LazyProc directly so that pointer arguments stay alive for the duration of the call. (#51074) The wrappers in the windows_eventlog_input operator routed every call through a SyscallProc interface, which hid the callee from the compiler and let the garbage collector free a buffer while the Windows API was still reading it. The interface is removed and tests mock the wrapper functions instead.

• processor/cumulative_to_delta: Fix histogram and exponential histogram reset recovery by updating prevPoint on reset detection (#50828)

• processor/delta_to_cumulative: Fix exponential histogram downscaling exceeding the bucket limit due to incorrect index rounding. (#50923)

• processor/drain: move units to singular form (#45270)

• processor/k8s_attributes: Reject an invalid exclude.pods[].name regex at config load instead of panicking at startup (#50957)

• processor/k8s_attributes: Emit the semantic convention deprecation warnings that were unreachable because they were evaluated before the extraction rules were populated (#51188)

• processor/metrics_transform: Preserve an absent histogram sum when aggregating datapoints instead of reporting a sum of 0 (#49379) When merging (exponential) histogram datapoints, the sum is now only kept when every merged datapoint carries one, matching the existing handling of min and max. Previously a merge that involved a datapoint without a sum produced HasSum=true with Sum=0, which is indistinguishable from a genuine zero sum.

• processor/resource_detection: Honor the configured timeout in the gcp detector when the metadata server is unreachable. (#50754) The gcp detector did not honor the configured timeout. It now checks the metadata server with a context that carries the deadline.

• processor/rolling_span_latency: move units to singular form (#45270)

• processor/signing: Fix int64 precision loss in canonical payload — timestamps and integer attributes are now serialized as quoted decimal strings instead of IEEE-754 doubles, preventing distinct records from producing identical signatures. (#50881)

• processor/tail_sampling: Reuse the original decision's threshold when rewriting tracestate for late-arriving spans of an already-sampled trace. (#50623)

• processor/transform: Ensure each batch of data receives a fresh cache when shared_cache is enabled (#51087)

• processor/transform: Fix exponential-to-explicit histogram conversion (#50737) The conversion now handles positive, negative, and zero buckets, includes the required overflow bucket, preserves data point metadata, and uses the exponential histogram mapping implementation for bucket boundaries.

• receiver/elasticsearch: Emit the elasticsearch.index.shards.size metric with aggregation: primary_shards in addition to aggregation: total (#48918) The index shards size metric only reported the total aggregation, unlike the other index metrics and its own documented attributes. It now also reports the primary_shards aggregation.

• receiver/fluent_forward: Cap decoded msgpack array entries at 1,000,000 and map options at 1,000 to avoid excessive memory allocation. (#49223)

• receiver/jaeger: Bound the Thrift decoder's max message size to the HTTP request body length to prevent a memory-amplification DoS. (#49218) A tiny Thrift-binary payload could declare a spans list with a huge element count, forcing the decoder to pre-allocate a large slice before failing to read the (absent) data. The Thrift /api/traces handler now caps the decoder's max message size at the number of bytes actually received, so a container that declares more elements than the payload can hold is rejected instead of being pre-allocated.

• receiver/k8s_events: move units to singular form (#51262)

• receiver/k8s_events: Keep the watcher in the Kubernetes events receiver retry when retrieving fresh resource version after a 410 Gone response but fails transiently. (#49962)

• receiver/k8s_objects: Keep the watcher in the Kubernetes objects receiver retry when retrieving fresh resource version after a 410 Gone response but fails transiently. (#49962)

• receiver/kafka: Stop reporting stale offset lag and current offset metrics for revoked Kafka partitions. (#36093) The otelcol_kafka_receiver_offset_lag and otelcol_kafka_receiver_current_offset metrics will now be reported using async gauges.

• receiver/kafka: Stop processing a fetched batch once the partition consumer is cancelled, instead of marking and dropping the rest of it (#50753) message_marking.on_error and message_marking.on_permanent_error no longer mark a record that was interrupted by the cancellation. At shutdown, records that were fetched but not delivered are redelivered on the next run instead of being committed.

• receiver/kafka: Apply metadata::retry::max and metadata::retry::backoff, which were silently ignored since the migration to franz-go. (#50116) metadata::retry::max governs retries for consumer group and offset commit/fetch requests. It does not apply to fetch requests, nor to the client's internal metadata refresh, which caps itself at 3 retries.

  metadata::retry::backoff is the minimum wait between retries; each successive retry doubles the wait, with jitter applied, capped at max(5s, backoff). Unlike max, it applies to every retry path, including fetches.

• receiver/kafka_metrics: Apply metadata::retry::max and metadata::retry::backoff, which were silently ignored since the migration to franz-go. (#50116) metadata::retry::max governs retries for the admin and offset fetch requests issued by the scrapers. It does not apply to the client's internal metadata refresh, which caps itself at 3 retries.

  metadata::retry::backoff is the minimum wait between retries; each successive retry doubles the wait, with jitter applied, capped at max(5s, backoff).

• receiver/kafka_metrics: Set the kafka.cluster.alias resource attribute to enabled when cluster_alias is defined. (#47573)

• receiver/macos_unified_logging: Stop re-emitting already sent records in live mode by resuming each poll from the last emitted record (#50625) Live mode previously recomputed --start as now - max_log_age on every poll and reset the poll interval whenever the re-read window returned records, so each record was emitted up to max_log_age / 100ms times. Text formats (default, syslog, compact) now also set the log record timestamp from the line. The json format has no per-line timestamp and is not de-duplicated.

• receiver/macos_unified_logging: Add predicate field aliases and support additional fields (#50626) Extended logType validation to include messageType alias and type validation to include eventType alias. Added support for eventMessage and signpostName predicate fields. Updated the README example to use the supported OR form for messageType/logType comparisons.

• receiver/mongodb: Fix replica set secondary connections never being established (#50622) Replica set secondaries were never discovered, so the receiver only ever collected from the instance it was pointed at. The receiver now connects to discovered secondaries, applying the configured TLS settings and an operation timeout to those connections.

• receiver/mysql: Disabling every metric fed by a query now also skips the query that fed it, instead of still running the query and discarding the result. (#50702)

• receiver/mysql: Check rows.Err() after iterating query results so a failed scrape is not reported as successful (#51125) Affected queries could return partial results without reporting an error, so a connection lost part-way through a scrape produced incomplete metrics that looked like a successful collection.

• receiver/netflow: Exclude netflowreceiver on solaris (#51328)

• receiver/oracledb: Fix top_query and top_procedure collection running before the configured collection interval has elapsed (#50888)

• receiver/oracledb: move units to singular form (#45270)

• receiver/otlp_json_file: Fix a panic and a zeroed accepted-data count by reading the span and metric counts before the data is handed to the next consumer (#50314) EndTracesOp/EndMetricsOp received SpanCount()/MetricCount() evaluated after ConsumeTraces/ConsumeMetrics had already returned. By that point a consumer that declares MutatesData may own the data - the batch processor's background goroutine moves the resource slice out and nils the source - so the receiver counted state it no longer owned. It reported 0 accepted spans or metric points, and could read a torn slice header and crash the collector. The logs path in the same file has captured its count before consuming since #29274.

• receiver/otlp_json_file: Attach the file attributes to profiles, as is already done for logs, metrics and traces. (#50562) Attributes such as log.file.name were silently dropped from profiles. They are now interned in the profiles dictionary and referenced from each profile, so options like include_file_name behave consistently across all four signals.

• receiver/postgresql: Fix incorrect top-query counter increments and query ranking caused by an undersized counter cache. (#51066) With the defaults (top_n_query: 200, max_rows_per_query: 1000), the cache held 4,000 entries for up to 10,000 counters, overflowing beyond roughly 400 statements when all ten counters were populated. Eviction within a collection caused full cumulative totals to be reported as interval increments on subsequent collections, also distorting which queries were selected as the top N. The cache is now sized from max_rows_per_query, independently of the reporting limit.

• receiver/postgresql: Prevent dropped top-query records and incorrect query-plan reuse when the same query ID appears under multiple databases or roles. (#51067) Rows from different databases or roles that shared a query ID previously used the same counter-cache entries. This could calculate deltas against unrelated counters and silently drop top-query records. Cache entries are now isolated by database, role, query ID, and counter column. Query plans are also cached separately for each database, role, and query ID.

• receiver/prometheus: Drop malformed classic histograms with decreasing bucket counts or overall count below largest bucket count instead of emitting invalid OTLP data. (#50348) Also applies when classic histograms are converted to NHCB via the convert_classic_histograms_to_nhcb scrape option.

• receiver/pulsar: Exclude pulsarreceiver on solaris (#51326)

• receiver/snmp: Fix a collector-crashing nil-pointer panic in the deferred client Close when a mid-scrape connection reset fails to reconnect. (#49703) The wrapper's Close now delegates to gosnmp's nil-safe, idempotent Close instead of dereferencing the connection field directly.

• receiver/splunk_hec: Extend the response write deadline while a request body is actively being read (#51391) Over HTTP/2 the server arms the WriteTimeout deadline when the handler starts and never extends it, so a large or slow request that is still actively transferring its body was reset (RST_STREAM) once WriteTimeout elapsed, even though it was making progress. The receiver now pushes the write deadline out as the request body is read, so WriteTimeout behaves as an idle timeout for the body-read phase rather than a hard cap on how long a request may take to arrive. The deadline is also reset once just before the data is handed to the pipeline so the downstream consumer starts with a full window; it is not extended while the consumer runs, so a consumer that blocks longer than WriteTimeout is still reset. When WriteTimeout is 0 (disabled) behavior is unchanged.

• receiver/splunk_hec: Set ObservedTimestamp on log records so downstream consumers have a valid timestamp when the event carries no time (#51391) Previously the receiver only set Timestamp, derived from the event's time field (events endpoint) or the ?time= query parameter (raw endpoint). When neither was present the timestamp defaulted to the zero value (Unix epoch) and ObservedTimestamp was never set, leaving records with no usable time. ObservedTimestamp is now stamped with the receiver's wall-clock time on both the events and raw paths, giving downstream components the standard fallback.

• receiver/ssh_check: Fix the configured timeout never being applied to the underlying SSH connection. (#50912)

• receiver/systemd: move units to singular form (#45270)

• receiver/tcp_check: move units to singular form (#45270)

API Changelog

🛑 Breaking changes 🛑

• all: Remove the pkg/pdatautil module. Its hash helpers moved to go.opentelemetry.io/collector/pdata/xpdata/xhash in opentelemetry-collector. (#51108) Replace imports of github.com/open-telemetry/opentelemetry-collector-contrib/pkg/pdatautil (Hash, Hash64, MapHash, ValueHash, and the With* options) with go.opentelemetry.io/collector/pdata/xpdata/xhash.

• exporter/elasticsearch: Remove the deprecated flush and num_workers configuration settings. (#42718) Migrate num_workers to sending_queue::num_consumers, flush::interval to sending_queue::batch::flush_timeout, and flush::bytes to sending_queue::batch::max_size with sending_queue::batch::sizer set to bytes. This also removes Config.NumWorkers, Config.Flush, and FlushSettings from the exported Go API.

• pkg/ottl: Unexport the implementations of the standard OTTL functions (#50927)

• pkg/ottl: Move profiles support into a separate pkg/ottl/contexts/xprofile module so it is excluded from OTTL's stability guarantees. (#50906) The ottlprofile and ottlprofilesample contexts moved to github.com/open-telemetry/opentelemetry-collector-contrib/pkg/ottl/contexts/xprofile, and the ProfileID converter and NewProfileIDFactory moved from ottlfuncs into that module's ottlfuncs package. ProfileID is no longer part of StandardFuncs and StandardConverters. Importers must update their import paths, depend on the new module, and merge the profiles converters (for example via xprofilefuncs.WithProfileConverters) to keep ProfileID available.

• pkg/ottl: Remove the unused Strp, Floatp, Intp, and Boolp helpers from the ottltest package. (#51037)

• pkg/ottl: Move the experimental lambda converters into the pkg/ottl/xottl module so they are excluded from OTTL's stability guarantees. (#51563) The All, Any, Filter, Find, MapEach, MapKeys, Reduce, and When converters and their New*Factory functions moved to github.com/open-telemetry/opentelemetry-collector-contrib/pkg/ottl/xottl/ottlfuncs. They are no longer part of StandardFuncs and StandardConverters; merge them (for example via xottlfuncs.WithExperimentalConverters) to keep them available.

• pkg/ottl: Unexport the ParserCollection Settings and ErrorMode fields in favor of read-only Settings() and ErrorMode() accessor methods (#51156)

• pkg/ottl: SliceGetter.Range now returns (bool, error); the boolean reports whether the underlying slice is non-nil. (#51494)

• pkg/ottl: Remove the ottltest package and its SetFeatureGateForTest function from the public API. (#47305)

• pkg/ottl: Remove the stable ottl.PanicDuplicateName and ottl.contexts.enableOTelColContext feature gates. (#44630, #46437)

• pkg/ottl: Unexport the argument structs of the standard OTTL functions (#51023)

• receiver/carbon: Move the protocol package to internal/protocol, removing it from the public API. (#51500)

🚀 New components 🚀

• pkg/ottl/xottl: Add the pkg/ottl/xottl module to hold OTTL's experimental Go API outside of OTTL's stability guarantees. (#51561)

💡 Enhancements 💡

• pkg/ottl: Add Experimental() to the Factory interface and a WithExperimental option so functions can declare whether they are covered by stability guarantees. (#50927)

• pkg/pdatatest: Add the /count collection operator to pmetricassert for resources, scopes, metrics, and datapoints. (#48472) <collection>/count takes exact, or min and/or max, and composes with /include. Pairing it with an exact <collection> list is a schema error.

• pkg/pdatatest: Add WithAttributeExists and WithAttributeRegex options to pmetricassert snapshot generation. (#49734)

• pkg/pdatatest: Add the /include collection operator to pmetricassert for resources, scopes, metrics, and datapoints. (#48472) Collections without an operator keep matching exactly. Within an item matched by /include, an omitted nested collection asserts nothing about it.

🧰 Bug fixes 🧰

• pkg/ottl: Fix SliceGetter.Get and literal slice folding so a coerced typed-nil slice (for example []string(nil) when elements are getters) is treated as nil instead of an empty non-nil slice. (#51494)
• pkg/ottl: Fix GetLiteralValues and GetScalarLiteralValues so they do not evaluate dynamic slices when checking for literal values. (#51438)
• pkg/ottl: Prevent SliceGetter.Get and SliceGetter.Range from panicking when constructed from nil. (#51174)
• processor/transform: Fix exponential-to-explicit histogram conversion (#50737) The conversion now handles positive, negative, and zero buckets, includes the required overflow bucket, preserves data point metadata, and uses the exponential histogram mapping implementation for bucket boundaries.

We are thrilled to welcome our first-time contributors to this project. Thank you for your contributions @hghukasyan, @Easonliuuuuu, @Rucha0901, @sneha-kr20, @sublimemonk, @SerhiiGoTolstoy, @krissiyan, @waheedi, @Lewis-E, @jochil, @abhinav1602, @suckatrash, @Mujib-Ahasan, @bachp, @insatomcat, @perebaj, @weliasz, @bingquanzhao, @Abhimanyu9988, @rahulgtt, @KR-Ravindra, @PySheriff, @chiliec, @MichalPawlus, @eKuG, @stelizhenets, @CedricConday, @weili-broadcom, @samuelvl, @abhishuraina ! 🎉