Agent

Known Issues

• GPU Monitoring can trigger a kernel panic on multi-GPU nodes with Hopper/Blackwell GPUs. The workaround is to disable parallel collection (gpu.parallel_collectors: false), downgrade to Agent v7.81.3 or upgrade to v7.82.1 when it becomes available.

Prelude

Released on: 2026-08-05

• Please refer to the 7.82.0 tag on integrations-core for the list of changes on the Core Checks

Upgrade Notes

• Automatic multi-line log detection (logs_config.auto_multi_line_detection) is now enabled by default. Multi-line log messages such as stack traces and JSON blobs are aggregated into a single log entry out of the box, instead of being split into separate entries. To restore the previous behavior, set logs_config.auto_multi_line_detection to false (or the environment variable DD_LOGS_CONFIG_AUTO_MULTI_LINE_DETECTION=false).
• Change default EVP track for AI usage to eudm-intake and disable AI usage agent desktop monitoring by default (the monitor stays in idle mode). AI usage agent's ai_usage_native_host.yaml configuration file is now regenerated from the packaged template on every Agent install and upgrade to ensure the default changes take effect. This is a deliberate short-term measure: it forces already-installed machines onto the updated defaults. Once the defaults are settled (expected within a few Agent releases), the installer will go back to preserving an existing ai_usage_native_host.yaml and only creating it when missing. Note that any customization made to the config file is discarded on every Agent install and upgrade, and must be re-applied afterwards.
• APM: Bump the default version of the datadog-apm-library-js package installed by the Datadog installer from major version 5 to major version 6, following the release of dd-trace-js v6.
• APM: Updates the default JS (Node.js) library used for Kubernetes auto-instrumentation (Cluster Agent admission controller) from major version 5 to major version 6.
• The legacy viper based configuration backend has been removed. The DD_CONF_NODETREEMODEL environment variable and the conf_nodetreemodel configuration setting no longer have any effect and can be removed from your configuration. The Agent now always uses the improved configuration implementation.
• serverless-init no longer forces DD_TRACE_PROPAGATION_STYLE=datadog during tracer auto-instrumentation. The tracer's own default (which includes W3C tracecontext and baggage in addition to datadog) now applies, and a customer-provided DD_TRACE_PROPAGATION_STYLE is respected. Applications that relied on serverless-init restricting propagation to datadog only should set DD_TRACE_PROPAGATION_STYLE=datadog explicitly.

New Features

• Private Action Runner: add the com.datadoghq.remoteaction.rshell.runRemediationCommand action. It behaves like runCommand but runs the restricted shell in remediation mode, which additionally permits file-target output redirections (>, >>, 2>, &>, &>>) and write-oriented builtins such as truncate, all confined to the configured allowed paths. The action is not enabled by default and must be explicitly added to the runner's actions allowlist.

• agent flare now includes diagnostic artifacts from the Agent Data Plane (ADP) process when data_plane.enabled is set to true. If ADP is unreachable at flare time, an UNREACHABLE.txt file containing the connection error is written to ADP's subdirectory and the rest of the flare completes normally.

• When infrastructure_mode is set to cloud_cost_only, the Agent adds an infra_mode:cloud_cost_only tag to metrics from selected integrations. Use integration.cloud_cost_only.tagged to list which checks receive the tag; when the list is empty (the default), all checks are tagged.

• Add a new dogstatsd_no_aggregation_pipeline_workers_count configuration option to control the number of parallel workers processing messages in the no-aggregation pipeline. Defaults to 1 to preserve existing behavior.

• Adds Datadog CSI driver telemetry to COAT, including volume publish and unpublish attempts as well as library resolution, download count and duration, cleanup, cache size, cached library count, and library volume link metrics.

• The Datadog OTLP connector now scales APM stats (hits, errors, duration) by the probabilistic head-based sampling weight carried in the W3C tracestate (th threshold and p power-of-two encodings). When upstream head-based sampling has dropped a fraction of traces, the computed trace metrics are scaled up to reflect the true traffic volume instead of only the sampled subset.

• Envoy Gateway AppSec protection can now run in sidecar mode over a Unix domain socket. Datadog injects the serviceextensions ext_proc container into Envoy Gateway data-plane pods, and Envoy Gateway communicates with it through an Envoy Gateway Backend.

  This behavior is selected by cluster_agent.appsec.injector.mode, which now defaults to sidecar. Envoy Gateway must have the Backend extension API enabled with extensionApis.enableBackend: true; if it is disabled, the cluster agent warns and does not change Envoy Gateway configuration.

  This is a behavior change for AppSec-enabled Envoy Gateway deployments: they now default to sidecar injection instead of external Service mode. To keep the previous behavior, set cluster_agent.appsec.injector.mode to external.

• Added an experimental telemetry error log forwarder. Disabled by default, the Agent forwards records logged at ERROR level or higher to the COAT intake so Datadog Engineers can aggregate Agent errors across customer organizations. The forwarder shares the agent telemetry transport, inheriting endpoint and compression settings from the agent telemetry configuration.

• On Windows, datadog-installer now honors the DD_AGENT_MAJOR_VERSION and DD_AGENT_MINOR_VERSION environment variables, matching the Linux and macOS install scripts.

• GPU: add the gpu.device.needs_recovery metric, which reports whether a GPU requires a recovery action (such as a reset or node reboot) as exposed by NVML's GPU recovery action field. The value is 0 when no action is needed and 1 otherwise, and the metric is tagged with recovery_action (none, reset, reboot, drain or drain_and_reset).

• The agent status command now includes a "Logs Agent Backpressure" section reporting per-component utilization of the logs pipeline and an overall HEALTHY/WARNING/SATURATED state, making it easier to see which pipeline stage is the bottleneck when logs are delayed.

• On macOS, the Agent can now be restarted directly from the web-based GUI (Agent Manager).

• New Agent Secret Backend: "windows.regkey"

• APM Single Step Instrumentation now supports setting tracer configuration options via the admission.datadoghq.com/apm-inject.tracer-configs pod annotation, the annotation-based equivalent of the apm_config.instrumentation.targets[].ddTraceConfigs option. The value is a JSON array of objects (for example [{"name":"DD_PROFILING_ENABLED","value":"true"}]) and each entry's name must start with the DD_ prefix.

• NetFlow: automatically detect and split Cisco FirePower/ASA bidirectional (NSEL) flow records into two unidirectional flow events. NSEL records carry initiator→responder and responder→initiator byte/packet counts in NFv9 fields 231/232/298/299; the agent now captures these fields via built-in mappings and emits a separate flow for each direction with correctly swapped src/dst addresses, ports, and interfaces. No user configuration is required.

Enhancement Notes

• Malformed ad.datadoghq.com/service.* and ad.datadoghq.com/endpoints.* annotations on Kubernetes services are now reported as Autodiscovery misconfiguration health events when the health platform is enabled. The issue is resolved automatically once the annotation is fixed.
• Adds Agent Data Plane packaging and launchd service support to macOS Agent packages.
• Adds Agent Data Plane packaging to Windows Agent MSI installs. ADP is supervised by dd-procmgr via processes.d/datadog-agent-data-plane.yaml, written by the fleet installer during postinst (same pattern as DDOT on Windows).
• Emit datadog.cluster_agent.kubernetes_actions.running when kuberenetes actions product is enabled and running.
• Podman receiver metrics collected via the Datadog Distribution of OpenTelemetry (DDOT) Collector are now correctly classified with origin opentelemetry_collector_podmanreceiver instead of falling back to opentelemetry_collector_unknown.
• Added the data_plane.stop_timeout configuration setting, which controls the graceful shutdown budget for the Agent Data Plane (ADP). When unset, it derives its value from aggregator_stop_timeout + forwarder_stop_timeout, so customizing either of those component timeouts now extends ADP's shutdown window in lockstep with the core Agent.
• On startup the Datadog Agent now validates the system-probe configuration against its schema and reports any violations through the Agent Health pipeline. Only the values the customer set in the configuration are validated. This can be disabled with health_platform.invalidsysprobeconfig_check.enabled.
• On Windows, the AI Usage Chrome Native Messaging host is now delivered as a fleet-managed Agent extension that is only installed when End User Device Monitoring is enabled (infrastructure_mode: end_user_device). It is no longer unconditionally installed by the MSI, and is skipped on Agent upgrades when End User Device Monitoring is disabled.
• APM: Added cardinality limits to client-side stats computation in the stats concentrator. These limits are no-op in the agent and are intended for use by the Go tracer.
• Agents are now built with Go 1.26.5.
• dd-procmgrd: Add write RPCs (Create, Start, Stop, ReloadConfig, GetConfig) for runtime control of managed processes.
• On Windows, the ddinjector system-probe telemetry now negotiates the counter contract version with the installed driver, reporting new crash and boot-recovery counters when the driver supports them and degrading gracefully against older drivers.
• Agent Cloud Authentication (delegated authentication) now discovers AWS credentials from additional sources on Agent builds that include EC2 metadata support. In addition to the previously supported static credentials (AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY) and EC2 instance metadata (IMDS), it now supports IAM Roles for Service Accounts (IRSA, via AWS_WEB_IDENTITY_TOKEN_FILE and AWS_ROLE_ARN) and ECS task role / EKS Pod Identity container credentials. The region used for the STS exchange follows the delegated_auth.aws.region setting, then AWS_REGION / AWS_DEFAULT_REGION, falling back to us-east-1. Shared config and profile loading (AWS_PROFILE, ~/.aws) is not used.
• When dogstatsd_flush_incomplete_buckets is enabled, the DogStatsD server now also finishes processing in-flight packets before sending metrics in the current (incomplete) aggregation window.
• Add a retry_on_failure configuration block to the Datadog serializer exporter (DDOT and OSS Datadog exporter). The defaults (initial interval 2s, multiplier 2, maximum interval 64s, maximum elapsed time 15m) and the default sending_queue settings (queue_size: 300, num_consumers: 1) and timeout: 20s mirror the historical forwarder defaults so behavior remains consistent when upgrading. Users can tune these to increase throughput; for example, raising num_consumers lets the exporter dispatch multiple HTTP requests to the Datadog intake in parallel.
• Add the datadog.serializerexporter.UseSyncForwarder feature gate (Alpha, disabled by default) to the Datadog serializer exporter. When enabled via --feature-gates=+datadog.serializerexporter.UseSyncForwarder, metric submission becomes synchronous: HTTP errors propagate back through ConsumeMetrics, are counted in the standard OpenTelemetry exporter metrics, and are retried by the exporter helper — fixing the silent-drop behavior where otelcol_exporter_send_failed_metric_points would not increase even on intake failures. Applies to both the embedded Datadog OpenTelemetry Collector (DDOT) and the OSS Datadog exporter (opentelemetry-collector-contrib). The Agent OTLP Ingestion pipeline is unaffected.
• GPU: improve GPU metric collection latency by collecting NVIDIA GPU collector data in parallel.
• Add a new trace_container_tag_promotion configuration option to the infraattributes processor in the Datadog Agent's embedded OpenTelemetry Collector (DDOT). When set to duplicate or rename, custom tags emitted by the processor (for example tags produced by podLabelsAsTags) are written with a datadog.container.tag. prefix so they are promoted into Datadog container tags and become visible in the Infrastructure tab of a span. The default value off preserves the previous behavior. Known Datadog and OpenTelemetry container semantic conventions, as well as service / env / version tags, are exempt and always written under their canonical key.
• OTLP ingest: The container_tag_promotion mode used by the infraattributes processor on the OTLP ingest traces pipeline is now configurable through otlp_config.traces.infra_attributes.container_tag_promotion (environment variable DD_OTLP_CONFIG_TRACES_INFRA_ATTRIBUTES_CONTAINER_TAG_PROMOTION). Accepted values are off, duplicate and rename. The default off matches the processor's own default, so promotion is opt-in and the previous behavior is preserved.
• Installer: The default install script now honors the DD_PROCESS_CONFIG_PROCESS_COLLECTION_ENABLED, DD_PROCESS_CONFIG_CONTAINER_COLLECTION_ENABLED, and DD_PROCESS_CONFIG_PROCESS_DISCOVERY_ENABLED environment variables (and their DD_PROCESS_AGENT_* aliases) to set the corresponding process_config collection toggles in datadog.yaml at install time
• Migrate to a local version of the Datadog connector rather than the implementation from opentelemetry-collector-contrib. The connector's configuration and behavior are currently unchanged.
• [netflow] Add the tos, dscp, and dscp_name fields to NetFlow flow payloads, exposing the IP Type of Service byte along with its decoded DSCP value and standard RFC name (for example EF or AF41). The dscp_name mapping now includes all IANA-registered codepoints.
• Reduce idle agent memory usage when Network Device Monitoring NetFlow is disabled (the default). The NetFlow flow aggregator is no longer allocated when the feature is off.
• Reduced per-sample allocations in the metrics aggregator, lowering CPU and garbage-collection overhead on Agents processing a high volume of metrics.
• Reduced the CPU and memory overhead of sensitive-data scrubbing when collecting Kubernetes resources for the Orchestrator Explorer. The command-line tokenizer regular expression is now compiled once instead of on every scrubbing call, and redundant per-token work was removed.
• Reduced memory allocations when converting Python strings to C strings across the RTLoader boundary.
• OTLP trace ingestion now reports otel.scope.name and otel.scope.version in addition to the deprecated otel.library.name and otel.library.version. Use the disable_otel_scope_convention feature gate to stop reporting the new keys.
• The OTel Agent --sync-delay flag (env DD_SYNC_DELAY) now defaults to 30s instead of 0. The OTel Agent will retry synchronizing its configuration from the core Agent for up to 30 seconds at startup before failing, which makes startup more resilient when the core Agent is not yet ready. This default can still be overridden via the flag or environment variable.
• The Private Action Runner now submits execution metrics through DogStatsD.
• Private Action Runner: restricted shell actions now combine Datadog execution policy allowlists with operator-configured command and path allowlists. Read-only and read-write allowed paths are handled separately, so remediation commands can write only to paths allowed by both policies.
• Persist low-priority transactions (such as retried metrics) to disk during a graceful Agent shutdown instead of dropping them.
• Added GPU attribution support for Kubernetes Dynamic Resource Allocation allocations reported by the kubelet PodResources API. The Agent maps NVIDIA DRA gpu-N device names to local GPU indexes.
• Reduced the CPU usage of log file scanning when tailing container logs.
• The image-size portion of the available-disk check performed before a container image SBOM scan now runs only for scans that export the image to a tarball on disk (the containerd default mode and the Docker collector), and is skipped for scans that read the image layers in place (CRI-O, and containerd with sbom.container_image.overlayfs_direct_scan or sbom.container_image.use_mount). The flat sbom.container_image.min_available_disk floor still applies to every scan, and its default is lowered from 1GB to 10MB.
• Bumped the Security Agent policies to v0.82.0
• serverless-init no longer overrides customer-provided CORECLR_ENABLE_PROFILING, CORECLR_PROFILER, CORECLR_PROFILER_PATH, or DD_DOTNET_TRACER_HOME values when auto-instrumenting a bundled .NET tracer. Any value already present in the environment is now respected.
• The OTel Datadog exporter will now emit otel.datadog_exporter.metrics.running.fargate{task_arn} for AWS ECS Fargate workloads so that each workload has its own metric. Host-based workloads continue to use otel.datadog_exporter.metrics.running{host} unchanged.
• APM: The minimal OTel-to-Datadog span conversion used for APM stats now preserves the raw W3C tracestate in the w3c.tracestate span tag, allowing downstream consumers to recover head-sampling probability.
• Update OpenTelemetry Collector dependencies to version 0.155.0. Notable upstream changes included in this update:
  • Datadog extension: Fixed tls.insecure_skip_verify being ignored.
• The discovery service map (discovery.service_map.enabled) now captures HTTP/2 traffic in addition to HTTP and TLS, so gRPC and other HTTP/2 service-to-service calls appear in the service map.
• Universal Service Monitoring (USM) now uses the direct consumer for HTTP monitoring by default on kernels >= 5.8.0, reducing the latency of HTTP event collection. On older kernels the batch consumer continues to be used. The previous behavior can be restored by setting service_monitoring_config.http.use_direct_consumer to false.
• On Windows, Universal Service Monitoring now derives the service, env, and version tags for IIS-hosted applications from the environment variables configured in applicationHost.config (application pool and aspNetCore environmentVariables) and in each application's web.config, in addition to appSettings and datadog.json. The resolution follows the .NET tracer's precedence so USM tags match what APM reports for the same application.
• On Windows, when the Agent is installed with the fleet OCI layout and the DDOT extension, the OpenTelemetry Collector is now supervised by dd-procmgr using a process definition under the Agent install layout, consistent with Linux. The datadog-otel-agent Windows service may still be registered for rollback, but the core Agent does not start it when that process definition file is present under processes.d and process_manager.enabled is true, so DDOT is not run twice while procmgr is on. The core Agent starts dd-procmgr-service when process_manager.enabled is true (set it to false to avoid starting the process manager if needed). The DDOT extension install hook writes processes.d and restarts dd-procmgr-service only under the same process_manager.enabled setting. Without the fleet processes.d DDOT definition, behavior is unchanged: the Agent still starts datadog-otel-agent when the collector is enabled.
• On Windows, when process_manager.enabled is true and the fleet installer writes processes.d/datadog-agent-action.yaml, the Private Action Runner is supervised by dd-procmgr instead of the legacy datadog-agent-action Windows service. When the processes.d definition is absent or process manager is disabled, the Agent continues to start PAR via SCM when private_action_runner.enabled is true.

Security Notes

• The trace-agent now scrubs sensitive values (passwords, tokens, API keys) from the command_line field of SSI injection-metadata telemetry payloads before forwarding them.

Bug Fixes

• Logs file tailer: directory entries returned by a configured path glob are now skipped instead of being opened as files. Previously the tailer attempted to open the directory and surfaced a misleading Access is denied error from the OS (on Windows in particular), which sent troubleshooting toward permissions. A non-wildcard path that resolves to a directory now returns a descriptive error explaining that a file or glob is required.
• Removed user_id field from AI usage event to avoid inconsistency between Chrome extension mode and desktop-monitor mode. user_id is taken care of by Datadog backend.
• [DBM] Bump go-sqllexer to v0.2.3 to fix a SQL normalization bug:
  • Preserve bracket-quoted T-SQL identifiers containing spaces (e.g. [Column With Spaces]) so they are no longer corrupted during normalization.
• Fixed a bug in the Go-native disk check (diskv2) where disk IO metrics (such as system.disk.read_time, system.disk.write_time, system.disk.read_time_pct, and system.disk.write_time_pct) were reported for every device regardless of the device_include and device_exclude settings. IO metrics are emitted only for devices whose partitions pass the configured device filters, matching the behavior of partition metrics and of the Python disk check.
• ECS daemon-scheduled tasks now emit the daemon_task_definition_arn tag instead of task_definition_arn, enabling correct entity resolution in the ECS Explorer.
• Fix a bug where the CELSelector field was not included in the Autodiscovery config digest. Configs with different CEL selectors were incorrectly treated as identical, which could cause the wrong workload filter rules to be applied.
• Fixed automatic multi-line Go stack trace aggregation for container-based log formats.
• CWS: Network packets of NAT-translated connections (for example source-port masquerading) are now correctly attributed to the owning process on the ingress path.
• Fix a bug where a Data Observability query action targeting a bare host shared by multiple database instances on different ports (e.g. two postgres instances on the same host) could drop every instance on that host from the remaining file-provider configuration, silently stopping normal DBM collection on the untargeted ports. Only the instance actually selected as the Data Observability check is now excluded.
• Fix Docker log parsing for TTY-mode containers when a single log line exceeds the 16KB Docker buffer size. Previously, the Agent retained the per-chunk timestamp prefix that Docker inserts at every 16KB boundary, causing those timestamps to appear inside the collected log content. The parser now strips the intermediate timestamps so the log line is reassembled correctly.
• Fix panic: runtime error: invalid memory address or nil pointer dereference when a DogStatsD distribution metric is submitted with a non-finite sample rate such as NaN (for example s.dist:1|d|@nan). The invalid sample rate is now treated as unsampled instead of corrupting the sketch's sample count.
• Fix stale metrics for locally-owned DatadogPodAutoscaler objects. The cluster agent cached the CRD object only on the first reconcile; status subresource writes (which do not bump .metadata.generation) were never reflected in the cached copy, causing the following metrics to report their initial values until the next spec or annotation change: datadog.cluster_agent.autoscaling.workload.status.desired.replicas, datadog.cluster_agent.autoscaling.workload.status.vertical.desired.container.cpu.request, datadog.cluster_agent.autoscaling.workload.status.vertical.desired.container.cpu.limit, datadog.cluster_agent.autoscaling.workload.status.vertical.desired.container.memory.request, datadog.cluster_agent.autoscaling.workload.status.vertical.desired.container.memory.limit, datadog.cluster_agent.autoscaling.workload.autoscaler_conditions. The controller now unconditionally refreshes the cached object on every reconcile.
• Fix error raising could not set '*.use_http' unknown key raised by the Fips proxy.
• Fixed the Cluster Agent kubeapiserver workloadmeta collector so that resources are now discovered on a non-preferred API group version when they are not served on the group's preferred version.
• Fix a potential goroutine hang in the OTLP logs exporter during graceful shutdown. The blocking channel send now respects context cancellation, allowing the exporter to stop cleanly instead of waiting indefinitely for the downstream logs pipeline.
• Fix the OTLP logs exporter sending corrupt data downstream when JSON marshaling of a log record fails. The malformed record is now dropped and the error is logged.
• Fix a bug in the OTel Agent where OTLP explicit-bucket histograms with bounds beyond the internal sketch's representable range could exhaust memory and crash the process, or — once the loop was bounded — record non-finite min, max, sum, and avg in the resulting sketch. Saturating bounds are now clamped to the largest representable finite bin.
• Fixed a rare process-agent panic when collecting metrics for a container with an empty or short (<12 character) ID.
• Fixed an issue where the Cluster Agent could fail to schedule Prometheus checks for Kubernetes Services newly annotated with prometheus.io/scrape until the Cluster Agent was restarted or the backing workload was scaled.
• Fixed a slow memory leak in the remote workloadmeta collector where a new context was created on every stream reconnection attempt without cancelling the previous one. This could cause unbounded growth in the number of orphaned contexts when the remote workloadmeta endpoint was repeatedly unreachable.
• Fix a bug where a SAP HANA instance targeted by a Data Observability query action could run as multiple duplicate check instances at once, causing duplicate database monitoring collection. The targeted instance is now correctly excluded from the remaining file-provider configuration.
• Fix a bug where editing an active Data Observability query action's monitor (e.g. changing its query count) could cause the previously excluded database instance to run as a duplicate check instance again, alongside its Data Observability check, causing duplicate database monitoring collection.
• Fixed several issues in the SBOM runtime usage enrichment ("package in use"). The HasSetSuidBit and RunningAsRoot properties are now reported as false for packages that are not in use (previously absent), so consumers can distinguish "not in use" from "unknown". A package's setuid observation is no longer cleared by a later access to one of its non-setuid files. An idle workload whose container image scan is slow is now enriched once the scan completes, instead of being skipped after a fixed number of retries.
• On Windows, fix Cross-Org Agent Telemetry (COAT) reporting for DDOT under dd-procmgr: runtime.agent_service_procmgr_configured{service:ddot} now checks InstallPath\processes.d where the installer writes the YAML, instead of ProgramData\Datadog\dd-procmgr\processes.d.
• On Windows, align the DDOT processes.d definition with the legacy datadog-otel-agent SCM service: run otel-agent.exe without CLI arguments and set only DD_OTELCOLLECTOR_INSTALLATION_METHOD=bare-metal. The collector resolves otel-config.yaml and datadog.yaml from ProgramData when paths are omitted, and reads fleet policies from the registry (with a stable managed-path fallback) when DD_FLEET_POLICIES_DIR is not set in the process environment. This fixes DDOT failing to start under dd-procmgr when the baked processes.d invocation or fleet policy handling did not match that SCM path (including an unsubstituted DD_FLEET_POLICIES_DIR placeholder).
• On Windows, stop baking DD_FLEET_POLICIES_DIR into Private Action Runner and Agent Data Plane processes.d definitions at install time. Those processes now resolve fleet policy location from the Windows registry via FleetConfigOverride, so fleet policy updates are not blocked by a stale install-time path.
• Exclude the workloadmeta process collector from the IoT Agent binary using the systemprobechecks build tag, reducing binary size.
• OTel Agent: Disable v3 series API shadow sampling, which is incompatible with the zlib compression the OTel Agent forces for the metrics intake.
• Fixed a bug where CPU core count fields (cpu_cores, cpu_logical_processors) were missing from OTel host metadata payloads in gateway topologies.
• Preserve OTLP attributes on Kubernetes manifests sent by the OTel Datadog exporter in a dedicated manifest attributes field.
• [oracle] Removed the oracle_client_lib_dir instance configuration option, which did not work as expected. The Oracle client library path must be set before runtime for the dynamic linker to properly link the library, so pointing to it from the check configuration is too late to take effect. Configure the client library location through the runtime linker instead, as described in the Oracle integration setup documentation: https://docs.datadoghq.com/integrations/oracle/?tab=linux#prerequisite
• Cluster Agent: fixes a bug where the Cluster Agent did not reset the dangling config and unscheduled check metrics when its internal state was reset.
• The container image SBOM collector no longer attempts to export an image to a tarball when the image's layers are not present in the containerd content store, as happens with remote snapshotters such as nydus. The export could never succeed for these images, and the scan was retried indefinitely, keeping the scan worker busy and increasing Agent memory usage. Such scans are now skipped instead of retried.
• Skip cloud provider network ID probes when cloud_provider_metadata is empty. Previously, GetNetworkID would probe GCE and EC2 metadata endpoints on every host metadata collection cycle (every 5-30 minutes) even when all cloud providers were disabled by configuration. The probes now short-circuit immediately, avoiding unnecessary work in on-premises environments.
• Reduce log noise in on-premises environments by changing the could not get network metadata message from INFO to DEBUG level. In environments where cloud provider metadata is disabled by configuration, this message was emitted at every metadata collection cycle (every 5-30 minutes), causing unnecessary log volume. The message is still available at DEBUG level for diagnostic purposes.
• Fix a potential panic and loss of latency data when discovery service map HTTP or HTTP/2 statistics are merged across collection cycles (for example when a client misses a cycle or multiple clients are registered).
• Disable v3beta metrics intake shadow payloads when zlib compression is used.
• On Windows, fixed an issue where a failed Agent upgrade could leave some files without their permissions restored after the MSI rolled back, which could prevent the Agent from starting.

Other Notes

• Agent Data Plane has been bumped to version 1.3.1. See the Agent Data Plane 1.3.1 release notes.
• The anomaly detection observer component no longer allocates memory or starts background goroutines when anomaly detection is disabled (the default). This reduces idle Agent PSS.
• Update libgcrypt to 1.12.2.

Datadog Cluster Agent

Prelude

Released on: 2026-08-05 Pinned to datadog-agent v7.82.0: CHANGELOG.

New Features

• The Cluster Agent's Prometheus HTTP Service Discovery provider supports an optional exclude_filter field per endpoint entry. The field accepts a CEL expression evaluated against each discovered target's host, port, and labels fields. Targets for which the expression returns true are skipped at collection time.

Enhancement Notes

• The Cluster Agent's leader election now uses a dedicated Kubernetes API server client with independently managed client-side rate limiting to be more resilient.
• Reduce scale-up stabilization windows for built-in DPA presets: Optimize Cost from 300s to 190s, Optimize Balance from 600s to 130s, and Optimize Performance from 900s to 70s. This allows faster scale-up response for workloads using autoscaling profiles.
• Cluster check stickiness is now enabled by default. The dispatcher biases check placement toward the runner where a check previously ran, reducing unnecessary check migrations. The behavior can be tuned or disabled via the following configuration options:
  • cluster_checks.stickiness_enabled — enable or disable stickiness (default: true)
  • cluster_checks.stickiness_factor — multiplier applied to check cost when computing the bias (default: 4.0)
  • cluster_checks.stickiness_upper_limit — maximum bias applied regardless of check cost (default: 1.0)
  • cluster_checks.stickiness_lower_limit — minimum bias applied when stickiness is enabled (default: 0.05)

Bug Fixes

• Fixed APM Single Step Instrumentation injecting the library twice when the admission webhook is reinvoked (for example on GKE Autopilot, where another mutating webhook triggers reinvocation). In CSI injection mode the pod has no init container, so the re-admission guard failed to detect that the pod was already instrumented and appended the injector to LD_PRELOAD a second time. The guard now also checks for the instrumentation volume, which is present in every injection mode.
• Fixed an issue where the Cluster Agent could associate a pod's detected languages with the wrong Deployment. The Cluster Agent now only attributes a pod's detected languages to a Deployment when the pod is owned by a ReplicaSet and the ReplicaSet derived from the pod name matches the owner ReplicaSet. Pods that are not owned by a ReplicaSet are no longer considered for Deployment-level language detection.
• Fix cluster_checks.nodes_reporting gauge drifting upward across leader elections. The metric is now correctly decremented when the cluster agent loses leadership and the node store is reset.
• Fix agent commands in DCA (listener should always be started)
• Fix permission in docker image when executing "/readsecret.sh" script with dd-agent user
• Fixed an issue in the KSM check where cluster-aggregate metrics (kubernetes_state.container.<cpu|memory>_requested.total, kubernetes_state.container.<cpu|memory|gpu|mig>_limit.total, and the initcontainer equivalents) reported incorrect cluster totals when the check ran with pod_collection_mode: node_kubelet. The aggregate metrics are now computed from a dedicated instance (running on the cluster-agent or a cluster-checks runner) in the new pod_collection_mode: cluster_aggregates_only mode, which watches all pods directly from the API server. To enable the fix, set the cluster_aggregates_enabled: true instance option on the node_kubelet and cluster_unassigned instances; those instances then suppress the affected accumulators, eliminating multi-source gauge collision at ingestion. Without that option the previous (colliding) behavior is unchanged, so it must be set alongside deploying the cluster_aggregates_only instance. The fix preserves the per-pod metric scaling benefit of node_kubelet mode while restoring correct cluster aggregates.
• Fix a bug in the orchestrator explorer check that led to trying to collect Kubernetes subresources under certain custom resource API groups.