Agent

Prelude

Released on: 2026-09-30

• Please refer to the 7.84.0 tag on integrations-core for the list of changes on the Core Checks

Upgrade Notes

• On Linux, the Datadog process manager (dd-procmgrd) is now a first-class service manager alongside systemd, upstart and sysvinit, and it is the default when the dd-procmgrd binary ships with the Agent. The OpenTelemetry Collector distribution is supervised through processes.d and the datadog-agent-ddot systemd unit is no longer installed.

  Each service manager now installs its own self-consistent set of units, so a host runs either the dd-procmgrd supervision path or the per-payload systemd units, never a mix of the two.

• Linux Agent packages no longer include the static archive /opt/datadog-agent/embedded/lib/libpcap.a. This affects users whose custom integrations or build scripts link directly against that archive. Before upgrading, check for references to this path and instead install the platform's libpcap development package or provide a separate libpcap build. The libpcap headers under /opt/datadog-agent/embedded/include remain available.

• The com.datadoghq.remoteaction.agent Private Action Runner bundle (Preview) has been renamed to com.datadoghq.remoteaction.datadogagent.

• The logs Agent now sends log payloads at a fixed higher concurrency by default instead of scaling the number of concurrent senders dynamically with intake latency (the previous "RTT fairness" behavior). This improves throughput for high-latency and high-volume workloads without manual tuning.

  The Agent now uses a static send concurrency of logs_config.pipelines x 10 (for example, 40 concurrent senders on a host with the default 4 pipelines).

  Connection impact — please review before upgrading: because the Agent now runs more senders concurrently by default, it may open more simultaneous connections to the logs intake (up to the static concurrency above). If you need to limit connections — for example behind a proxy or on a connection-constrained network — set logs_config.batch_max_concurrent_send to a lower value (1 keeps a single sender per pipeline).

New Features

• Data security PostgreSQL scans can now connect over TLS. The instance option ssl accepts disable, allow, prefer and require. The verify-ca and verify-full modes are not supported yet, as server certificates are not validated.
• Data security PostgreSQL scans now support the verify-ca and verify-full SSL modes, and send a client certificate and key on TLS connections when ssl_cert and ssl_key are set.
• On Linux and macOS, the Agent now reports an Agent Health issue when the log file tailer loses log data because a file was rotated before the tailer finished reading it. The issue is reported once per host and covers every affected log source, reporting the total bytes lost and rotations involved over the last 24 hours together with a breakdown per source and service. Loss recorded before an Agent restart is not carried across the restart.
• On macOS, the thermal check now reports hardware temperatures and the system thermal pressure level. It submits system.thermal.temperature.cpu, system.thermal.temperature.gpu, system.thermal.temperature.ssd and system.thermal.temperature.battery in degrees Celsius, along with system.thermal.pressure_level (0 nominal, 1 moderate, 2 heavy, 3 trapping, 4 sleeping) tagged with the level name. Both Apple Silicon and Intel Macs are supported. Sensors that the host does not expose are omitted rather than reported as zero, so the set of available metrics varies by hardware model.
• OTLP ingestion: Adds a new otlp_config.metrics.infra_attributes.as_tags option. When enabled, custom tagger-derived tags (for example, tags configured via kubernetesResourcesLabelsAsTags/kubernetesResourcesAnnotationsAsTags) are promoted so they survive the metrics translator's allowlist and are emitted as metric tags for OTLP metrics ingested directly by the Agent. Without it, custom tags that are not known Datadog or OpenTelemetry conventions are dropped. Default behavior is unchanged.
• DDOT: The infraattributes processor now supports a new metrics_attributes_as_tags option for the metrics pipeline. When enabled, custom tagger-derived tags (for example, tags configured via kubernetesResourcesLabelsAsTags/kubernetesResourcesAnnotationsAsTags) are promoted so they survive the metrics translator's allowlist and are emitted as metric tags. Default behavior is unchanged.
• Rshell privileged helper process is introduced and allows specific, allow-listed commands to run with elevated privileges. This is enabled with private_action_runner: enabled: true restricted_shell: privileged: enabled: true it requires seccomp and landlock and is linux-only.
• Added the device_tags_source option to the SNMP check, controlling where the device tags on metrics come from. resource (default) sends only the device resource tag and lets the backend attach the device tags from the metadata payload. agent has the Agent attach the device tags to every metric and omits the resource tag, so no backend enrichment happens. both sends the device tags and the resource tag. Has no effect when collect_device_metadata is disabled.

Enhancement Notes

• The Datadog installer's setup script now accepts a DD_LOG_LEVEL environment variable to set the log_level option in datadog.yaml at install time. On Windows, the Agent MSI installer also exposes a DD_LOG_LEVEL property and passes it through to the Datadog installer.
• Network Path Synthetic tests now report an NDM namespace on emitted paths. By default the Agent uses its configured network_devices.namespace (matching the network_path integration); individual tests can override it via a namespace field in their configuration. Previously synthetic paths were always emitted with an empty namespace.
• Adds data_plane.preflight_mode_duration, which sets how long the Agent Data Plane (ADP) pre-flight leaves ADP running. It defaults to 90s and is clamped to that as a minimum, so it can only extend the window: a shorter one stops ADP while its startup is still in progress and reports a healthy host as a failure. It is intended for benchmarking harnesses that need ADP resident for the whole of a fixed-length run; on a real host the setting to reach for remains data_plane.preflight_mode.
• APM : The trace-agent EVP proxy now forwards the DD-EVP-ORIGIN and DD-EVP-ORIGIN-VERSION request headers to Event Platform intake. Previously these headers were stripped by the proxy allowlist, so server SDK metadata (SDK name and version) was lost at the Agent even when the SDK sent it. Values are forwarded unchanged; the Agent does not synthesize defaults for requests that omit them.
• APM : Use the OpenTelemetry url.template attribute in HTTP client resource names when available, while retaining method-only resource names as the fallback.
• APM: Add apm_config.traces_send_to_main_endpoint (DD_APM_TRACES_SEND_TO_MAIN_ENDPOINT, default true). This setting is for internal use, and we expect to remove it in a future version. When set to false, the trace-agent trace and stats writers stop sending to the main endpoint (api_key + apm_config.apm_dd_url) and forward traces and APM stats only to apm_config.additional_endpoints. The main endpoint's API key is still used by the other trace-agent proxies, and the configuration is rejected when no additional endpoint would remain. This is the traces/stats counterpart of apm_config.profiling_send_to_main_endpoint.
• APM : Add support for installing APM injection on AMD64 systems that also run 32-bit binaries. 32-bit executables run without injection, while supported 64-bit executables continue to be instrumented.
• Agents are now built with Go 1.26.6.
• Agents are now built with Go 1.26.7.
• The DDOT configuration converter now reuses a user-defined pprof, zpages, health_check or ddflare extension that was declared under extensions but not wired into service.extensions, instead of adding a default <name>/dd-autoconfigured copy. This matches the behavior already in place for the datadog and dogtel extensions, so a user's custom extension configuration is honored even when they forget to add it to the service's extension list.
• The Datadog Distribution of the OpenTelemetry Collector (DDOT) now sends series metrics to Datadog using the v3 metrics intake, reducing metrics egress bandwidth. This aligns DDOT with the core Agent: series follow the use_v3_api.series.enabled setting, whose default (datadog_only) uses the v3 intake for Datadog destinations while other destinations continue to use the v2 intake. To keep using the v2 intake, set use_v3_api.series.enabled to false (or the DD_USE_V3_API_SERIES_ENABLED environment variable).
• DogStatsD diagnostic commands (dogstatsd-stats, dogstatsd-capture, dogstatsd-replay, dogstatsd top, and dogstatsd dump-contexts) now print a clear error message and exit when invoked against the Core Agent while the Agent Data Plane is configured to handle DogStatsD traffic. In that mode the Core Agent's DogStatsD pipeline is dormant and the commands would silently produce empty or misleading results. Use the equivalent commands through the agent-data-plane binary instead.
• Added retry-with-backoff to startup of the External Metrics Provider. The Cluster Agent now retries the external metrics server setup on transient APIServer failures instead of failing on the first attempt.
• gpum: Add gpu.legacy_sm_active config toggle. When enabled on GPM-capable NVIDIA GPUs, gpu.sm_active reports the GPM SM utilization value while gpu.sm_utilization remains available.
• gpu: add gpu_nvlink_capable and gpu_nvlink_version tags to GPU metrics, to allow filtering GPUs that have NVLink enabled and the version supported.
• Health Platform: the health_platform.issues_detected telemetry counter is now also tagged with severity, in addition to the existing issue_type tag. This allows filtering or grouping detected health issues by their severity level.
• Tag Kubernetes Job events emitted by the kubernetes_apiserver check with kube_cronjob when the Job's name matches the pattern generated by a CronJob, similar to the tagging already applied to Pod events.
• agent status now reports why a file matched by a log configuration is not being tailed when its fingerprint cannot be used, under the log source that matched it. The most common cause is a file that holds less data than logs_config.fingerprint_config.count, for instance right after a log rotation replaced it with a smaller file, and the message names the setting to change. This previously showed up only as a shortfall in the N files tailed out of M files matching line, with no explanation.
• The Logs Agent now logs a warning when a file matched by a log configuration is not tailed because its fingerprint cannot be used, which previously happened without any log line at any level. The most common cause is a file that holds less data than logs_config.fingerprint_config.count, for instance right after a log rotation replaced it with a smaller file. The warning reports the path, the current size of the file and how much data fingerprinting requires, and distinguishes a file that is simply too short from a file whose fingerprint could not be computed at all. It is emitted once when the file starts being skipped rather than once per check, and a closing message is logged when it stops being skipped, whether the file started being tailed again or was never tailed at all, so the collection gap can be measured from the Agent log alone.
• The logon duration event on Windows now breaks each boot Group Policy pass down into the individual client-side extension invocations that ran during it, reporting the offset, duration, and outcome of each along with the Group Policy objects that fed it. The breakdown appears as a new group_policy_details block in the event's custom attributes, split into computer and user arrays.
• Notable Events on macOS now reports the cause of the previous shutdown on Apple silicon: the Agent classifies the power management unit's boot-fault record after each boot and emits a System shutdown fault event when the previous shutdown was caused by a power fault, a processor crash signal, a watchdog timeout, a hardware fault or a thermal fault. The event reports the fault classification and the underlying fault tokens, and is emitted at most once per boot.
• The Private Action Runner can now validate MongoDB connections
• The Network Configuration Management (NCM) check now emits a datadog.ncm.check_failure metric, tagged with the failure reason, whenever the check encounters an error.
• Network Config Management (NCM) local configuration store is now bounded by the network_devices.config_management.store.min_configs_per_device, network_devices.config_management.store.max_configs_per_device, and network_devices.config_management.store.max_raw_config_store_bytes configurations. Once these limits are exceeded, the least-recently-used configs are evicted. min_configs_per_device and max_configs_per_device is floored at 2. min_configs_per_device and max_configs_per_device are hard limits and will be enforced even if the size of the database file is greater than or less than max_raw_config_store_bytes. Default values are as follows: min_configs_per_device = 3, max_configs_per_device = 50, max_raw_config_store_bytes = 2000000000. Updates to store configurations become active upon agent restart.
• The Datadog Distribution of OpenTelemetry (DDOT) Collector now compresses all telemetry signals with zstd, providing a consistent compression algorithm across metrics, logs, and traces. Metrics and logs default to zstd level 3 and the level is configurable through the serializer_zstd_compressor_level and logs_config.zstd_compression_level settings. Previously, metrics were compressed with zlib and traces with gzip.
• OTLP ingest and DDOT: Logs received through the OTLP receiver now map the instrumentation scope name and version to otel.scope.name and otel.scope.version. Incoming otel.library.name/otel.library.version attributes (the deprecated OpenTelemetry predecessors) are remapped to the canonical otel.scope.* keys.
• Private Action Runner: the api_key_only_enrollment setting now defaults to true. Runners now enroll using only an API key by default, without requiring an application key. Your API key need to have "Private Action Runner" scoped enabled.
• Add opt-in PAR split mode support to the containerized Agent.
• Enables PAR action execution on-demand to significantly reduce idle memory usage. This is opt-in behind private_action_runner.split_enabled and is currently available on Linux host deployments.
• Private Action Runner: add the private_action_runner.restricted_shell.disable_detailed_telemetry setting. When set to true, it suppresses the raw command text and effective sandbox configuration that rshell attaches to its "run" telemetry span for each runCommand/runRemediationCommand invocation. The setting defaults to false; other rshell telemetry (exit code, timing, command counts) is unaffected.
• Adds support for on-demand PAR action execution to Windows host deployments.
• Data Observability: Extended the queryactions component to support SQL Server in addition to PostgreSQL. The component now schedules data_observability.queries for sqlserver check instances and correctly resolves Azure SQL Database instances by requiring both host and database equality, preventing cross-database payload injection.
• Autodiscovery now reports the check configuration keys it ignores when several annotation or label formats are set on the same entity. Only the format with the highest priority is applied (checks, then check_names with init_configs and instances, then the legacy service-discovery.datadoghq.com prefix), and the others used to be discarded silently. The ignored keys are now listed in the Autodiscovery section of the agent status output.
• The SSI injection-metadata telemetry payload now supports an optional, free-form metadata field for carrying result_class-dependent data.
• Bumped the Security Agent policies to v0.84.0
• APM: Reduce allocations when decoding v0.4 traces by interning strings directly from the incoming payload instead of allocating a string for every value already present in the string table.
• APM: Reserve room for the trailing end-of-body read when buffering an incoming trace payload, avoiding a reallocation and copy of the buffer.
• A small sample of sketch metric flushes (0.1% by default) is now additionally sent to a v3beta metrics intake endpoint to validate the upcoming v3 metrics protocol. Shadow traffic is only sent for agents configured against the datadoghq.com (US1) site. To opt out, set serializer_experimental_use_v3_api.sketches.shadow_sample_rate to 0.
• The Agent now logs a warning instead of an informational message when it finds a Docker, CRI or PodResources socket that exists but cannot be reached. On Unix this condition is only reported when opening the socket fails with a permission error, so it means the Agent user lacks access to the socket. Because these messages are emitted while the configuration is still loading, they were previously discarded on Agents running with log_level set to warn or above.

Deprecation Notes

• Remove system-probe module-restart CLI command.

Security Notes

• Data security PostgreSQL scans now open connections as read-only, as a first layer of protection against accidental writes.
• The Datadog flare extension (ddflare) in DDOT no longer serves its endpoint, by default https://localhost:7777, without authentication.
• Update agent-payload to v5.0.209 to remove the legacy zstd_0 dependency.

Bug Fixes

• Autodiscovery: a negative index in the %%port_<index>%% template variable (for example %%port_-1%%) no longer crashes the Agent. Negative indexes are now resolved Python-style, -1 being the last port, -2 the second to last, and so on, wrapping around the list of ports so that indexes beyond the number of ports still resolve.
• The file-based secret backend now rejects directory paths passed as a secret name. On AIX, reading a directory with os.ReadFile returns raw directory bytes instead of an error, which could cause a directory's contents to be returned as a secret value.
• APM: Raise the messagepack decoder allocation limit for the span meta_struct field to 10MiB, so large meta_struct entries (as written by LLM Observability) are no longer rejected by the package-wide 500,000 element limit. Other span fields keep the lower limit.
• APM peer-tag aggregation now refreshes derived tag keys when Remote Configuration changes semantic registry mappings without changing the producer-declared content hash.
• Add the origin product source to distribution metrics originating from checks.
• Clarify the Process Component status when Service Discovery is enabled without Live Process collection. The enabled checks now report service_discovery instead of process and rtprocess in this configuration.
• CSM Misconfigurations: a Kubernetes configuration file the Agent may not read is now reported with its content left out. An unreadable kubelet kubeconfig used to parse into an empty one, matching a kubeconfig that really names no cluster, and the managed environment detection concluded from it that an EKS, GKE or AKS node was unmanaged. The detection now sees an absent kubeconfig, and the payload says as much. The ownership and permissions of an unreadable file are reported too, so the benchmarks that check them see the real mode.
• CSM Misconfigurations: the Kubernetes node configuration collected for the CIS Kubernetes benchmarks now assumes the KubeletConfiguration defaults for a kubelet started with --config, and folds in the drop-ins of --config-dir. Kubernetes keeps the historical command line defaults of --read-only-port, --anonymous-auth and --authorization-mode for a kubelet started on flags alone, and the Agent applied them in both cases. A node whose configuration file left those settings out was therefore reported with a read-only port on 10255, anonymous authentication enabled and an AlwaysAllow authorization mode, while the kubelet was really running with the read-only port disabled, anonymous authentication disabled and Webhook authorization. OpenShift and kubeadm both leave readOnlyPort out of their rendered configuration, so their nodes failed the "kubelet read-only port should be disabled" rule with the port closed.
• CWS: fix a regression where the process context updates carried by an event (setuid, setgid, capset, login UID and IMDS security credentials) were applied before the event was evaluated, preventing rules from matching on the process state that preceded the event.
• Data Observability query actions now match PostgreSQL checks by their resolved database identifier. Queries now run when the identifier uses an agent hostname override or a database_identifier template.
• Fixed DDOT being unable to reach the core Agent in containerized deployments that do not pass --core-config (Docker, ECS, and ECS Fargate). Without a core config path the connection failed with x509: certificate signed by unknown authority. The OTel Agent now falls back to the default datadog.yaml location when it exists and the collector is not running in standalone mode.
• Fixed a bug in the trace-agent DogStatsD proxy endpoints (/dogstatsd/v1/proxy and /dogstatsd/v2/proxy) where a request body was split into all of its newline-separated payloads at once, so a body containing many newlines used several times its own size in memory. The body is now scanned one payload at a time, empty payloads are skipped, and the number of payloads relayed per request is capped.
• The Agent no longer shows the content of check config files on its local expvar page. This includes the configuration of JMX checks. The content is still sent to Datadog and still included in the flare.
• The metric filter list (metric_filterlist) now matches on the normalized metric name instead of the raw submitted name. Metric names are normalized by the Datadog intake on ingest, so a metric submitted as my metric-name is stored and displayed as my_metric_name. Previously a filter list entry using the normalized name that users see in Datadog would fail to match such a metric, and the metric was submitted anyway. Filter list entries themselves are matched as written, so they should be the metric name as it appears in Datadog.
• Kubernetes orchestrator: Prevent unchanged clusters from being repeatedly reported as updated when the node listing order changes.
• CWS: Network events are now attributed to the correct process when the same address and port are reused across different network namespaces.
• Fix the default query_timeout for the Oracle check from 20,000 seconds to 20 seconds.
• Fixed a bug in container image reporting where image references whose registry host included a port (for example myregistry.local:5000/foo/bar:1.2.3) were split on the first colon instead of the tag separator. This caused the image_name and image_tag tags, as well as the values shown in the Container Images view, to be incorrect for such images. The repo and tag are now parsed using the last colon following the last slash, matching standard image reference rules. Relatedly, the registry of images whose reference has a registry host followed by a single path component (for example localhost:5000/service or registry.k8s.io/pause) is now reported in the Container Images view instead of being left empty.
• Fixes Agent installs and upgrades failing to create any systemd unit files on hosts whose kernel does not support ambient capabilities (kernel older than 4.3). On those hosts the installer selects the -nocap systemd unit templates, but those templates were never compiled into the installer binary, so unit generation failed with failed to write stable units: open tmpl/gen/debrpm-nocap/datadog-agent.service: file does not exist and the host was left with no Datadog units at all. Because the package manager scriptlet ignores this failure, the install appeared to succeed while systemctl start datadog-agent reported Unit not found. This affected both the classic DEB/RPM install path and Fleet Automation remote upgrades and configuration experiments, which use the equivalent oci-nocap templates.
• Fix container log corruption when partial records from stdout and stderr are interleaved. The Agent now reconstructs partial CRI and Docker JSON-file records independently for each stream.
• Cluster Agent (KSM check): fix a collision when collecting custom resource metrics for two custom resources that share the same Kind and plural name but belong to different API groups (for example Project in both artifactory.example.com and sonarqube.example.com). Previously the resources shared a single API client, causing repeated Unexpected watch event object gvk errors and mixed, incorrect metric counts. Custom resource clients are now keyed by their fully-qualified GroupVersionResource.
• NCM check frequency will no longer default to a negative number; config values expressed as integers instead of durations (e.g. "5" instead of "30s" or "10m") will be parsed as seconds instead of nanoseconds.
• DDOT: fix DD_SITE being ignored when api.site is absent from the Datadog exporter config. Telemetry was silently sent to api.datadoghq.com instead of the site configured via DD_SITE or datadog.site.
• Fix an issue on Windows where uninstalling the Agent could fail if the configuration directory (C:\ProgramData\Datadog by default) had already been removed before running the uninstall.
• On Windows, the wlan check no longer panics on hosts where wlanapi.dll is unavailable. The library ships with the Wireless-Networking feature, which is not installed by default on Windows Server. Such hosts are now reported as having no active Wi-Fi interface.
• Fix an issue where the Agent logged a spurious error about Workloadmeta collectors not being ready on every startup in environments where no container runtime or orchestrator is detected, such as container sidecars.
• Flare archive filenames now include a process ID and counter suffix (datadog-agent-<timestamp>-<pid>-<counter>-<loglevel>.zip), preventing two archives created by the same Agent process at the same second from overwriting each other.
• Fleet Installer: Fix an issue where the installer's telemetry client repeatedly logged failed to send telemetry payload warnings with 404 Not Found errors on GovCloud sites (ddog-gov.com), since no instrumentation telemetry intake exists there. Telemetry is now disabled outright for GovCloud sites, matching the existing behavior of the Agent's own resident telemetry.
• HA Agent: a Remote Config document on the HA_AGENT product that belongs to the comp/workloadbalancing component (identified by an explicit type field) is now skipped instead of being processed as an invalid HA Agent document. If every document in an update batch turns out to belong to comp/workloadbalancing, HA Agent resets its state to Unknown rather than keeping a stale Active/Standby state.
• Fix Private Action Runner startup delays when signing keys are already available from Remote Config.
• Fix an Agent IPC client socket leak when a local service accepts a TLS connection but never completes the handshake.
• KSM core check: Fixed a bug where the kubernetes_state.configmap.count metric could stop being reported. ConfigMaps are collected using a metadata-only Kubernetes client, and the conversion of watch events into ConfigMap objects was dropping annotations, including the k8s.io/initial-events-end bookmark annotation used by newer versions of client-go's watch-list feature to signal that the initial list has finished syncing. Without that signal, the underlying reflector would wait indefinitely and the metric would never be reported.
• Windows: Fix the logon duration event reporting boot_timeline and group_policy_details entries out of chronological order. A milestone that ran while the machine sat at the login screen, such as Computer Group Policy on a domain-joined host, could render after milestones that actually followed it.
• Add new bind_host configuration option to TCP and UDP log listeners.
• Normalize Windows device tags to use forward-slash paths, preventing duplicate disk and IO metric device tags.
• Fixed a standalone DDOT (DD_OTEL_STANDALONE=true) failing to start when configured without a Datadog exporter, for example when the standalone DDOT is used to forward telemetry to a separate gateway layer via an OTLP exporter instead.
• Agent OTLP Ingest no longer attaches the OpenTelemetry debugexporter by default. The debugexporter is now attached only when the otlp_config.debug section is explicitly configured. Declaring the section without a verbosity uses the default verbosity (basic), while setting otlp_config.debug.verbosity (or DD_OTLP_CONFIG_DEBUG_VERBOSITY) to basic, normal, or detailed selects the verbosity. Setting it to none leaves the exporter detached.
• The deprecated process_config.enabled setting no longer overrides process_config.container_collection.enabled and process_config.process_collection.enabled when those are configured directly, whether through the configuration file, an environment variable, or any higher-precedence source. Previously, setting DD_PROCESS_CONFIG_ENABLED=false together with DD_PROCESS_CONFIG_CONTAINER_COLLECTION_ENABLED=false left container collection enabled. process_config.enabled still applies to whichever of the two settings is left unset, so configurations that only use the deprecated setting are unaffected.
• Autodiscovery now retries check configurations that failed secret resolution when secret_refresh_interval is enabled, allowing checks to recover after a transient secret backend outage without restarting the Agent.
• Fix a crash of the Agent process when a Python check raises an exception whose message contains a Unicode lone surrogate.
• Runtime usage enrichment keeps container image SBOM components that share a name and a version. Trivy reports one component per install location, so a library version pinned by two lockfiles appears twice, and the merge kept only the first, leaving the dependency graph referencing a component the payload had lost. Components are deduplicated by their CycloneDX bom-ref, which is what identifies one.
• Fixed container image SBOMs being reported as in use after their last container had stopped. On Kubernetes nodes, an image that had run a container once kept that flag until the Agent was restarted.
• The runtime usage properties merged onto container image SBOMs (LastSeenRunning, HasSetSuidBit and RunningAsRoot) now reach OS packages alone. The system-probe report is matched to components by name and version, so a language package sharing an OS package's name and version took the timestamp and flags of the OS package that had run. Components whose purl places them outside the dpkg, rpm and apk databases are left out of the match.
• The runtime usage properties merged onto container image SBOMs (LastSeenRunning, HasSetSuidBit and RunningAsRoot) are now set on OS packages alone. The runtime scanner reads the dpkg, rpm and apk databases, so language packages (npm, pypi, golang and others), the image's operating-system component and the per-lockfile application components stay out of its scope. The absence of a property marks a component out of scope, where a LastSeenRunning of 0 states that the package was watched and found idle.
• Fixed sbom.container_image.use_spread_refresher never refreshing container image SBOMs on hosts with fewer than ten images, and refreshing them more slowly than periodic_refresh_seconds on other hosts.
• Fixed the Agent crashing when the stored SBOM of a container image could not be uncompressed. The image is now skipped instead.
• ECS Fargate: skip EC2 IMDS instance-type lookups. The Agent no longer periodically queries 169.254.169.254/latest/meta-data/instance-type on Fargate (where IMDS is unavailable), which removes recurring INFO log noise in CloudWatch.
• SNMP: Fix the detection of profiles using the legacy Python metric syntax. The detection result is now cached along with the profiles, so every check instance is consistently handed over to the Python loader instead of only the first instance to be configured. Previously the remaining instances silently kept using the Core loader.
• SNMP: The error reported when a legacy profile forces the fallback to the Python loader now names the profiles using the legacy syntax.
• Stop the Agent from attempting to reach a Kubelet when running as a Cluster Checks Runner. Cluster Checks Runners are Deployment replicas, not DaemonSets, so they never have a locally-reachable Kubelet, since a CCR doesn't correspond to any one node. This removes the recurring Impossible to reach Kubelet through HTTPS warning logged by Cluster Checks Runner pods. There is no change in behavior on the node Agent or Cluster Agent.
• Auto multi-line detection no longer concatenates consecutive IIS W3C extended-format access log entries. A client IPv4 address next to the leading timestamp was being scored as part of that timestamp, which dropped the match to the detection threshold so each single-line record was treated as a continuation of the previous one. IPv4 addresses are now recognized as their own token and no longer interfere with timestamp detection.
• Fixed a crash in the trace-agent when processing a v0.7 payload whose trace chunk omits the tags field and whose spans carry a _dd.p.dm tag. Promoting the decision maker to the chunk level no longer writes to an uninitialized map.
• APM: Fix several trace-agent debug/error log messages that printed incorrect info due to format-string bugs.

Other Notes

• Agent Data Plane has been bumped to version 1.6.0. See the Agent Data Plane 1.6.0 release notes.
• Agent Data Plane has been bumped to version 1.6.1. See the Agent Data Plane 1.6.1 release notes.
• Added origin mapping for the Cisco Catalyst Center integration.
• Extended the delegated authentication (Workload Identity Federation) component so it can write a resolved API key into a map- or list-shaped additional_endpoints configuration value, replacing a placeholder DELA(...) directive. This is internal foundation work; no Agent subsystem reads DELA(...) directives yet, so this does not enable any user-facing behavior on its own.
• Added origin mapping for the Kueue and External Secrets integrations.
• The battery and wlan check configurations are no longer shipped in the Linux and AIX packages. Both checks can only collect on macOS and Windows, so on other platforms their configuration only mattered when infrastructure_mode was set to end_user_device, where it scheduled a check that could never report data. macOS and Windows packages are unchanged.
• Add metric origin mapping for the thermal integration.

Datadog Cluster Agent

Prelude

Released on: 2026-09-30 Pinned to datadog-agent v7.84.0: CHANGELOG.

Upgrade Notes

• DD_INSTRUMENTATION_INSTALL_TYPE=k8s_single_step and SSI defaults (for example DD_TRACE_ENABLED) apply only when a target or policy matches the pod, not merely because the namespace could match some rule. Pods that only have library annotations and do not match a target or policy use k8s_lib_injection.

New Features

• DatadogInstrumentation checks and logs configurations can now target Strimzi StrimziPodSet workloads.
• Add AppSec injection support for GKE managed Gateways (EXTERNAL mode only). The Cluster Agent now detects GKE managed Gateways whose spec.gatewayClassName is in an allowlist of external-managed GatewayClass names (gke-l7-global-external-managed, gke-l7-regional-external-managed) and creates one GCPTrafficExtension (networking.gke.io/v1) per Gateway to route edge traffic through a user-deployed Datadog AppSec callout service. SIDECAR mode is not supported because managed GKE has no in-cluster Envoy data plane. The callout Deployment, Service, and HealthCheckPolicy must be deployed by the user following the public GKE service-extensions documentation. Cluster-agent RBAC for gcptrafficextensions.networking.gke.io (get/list/watch/create/delete) is required. The GatewayClass allowlist is configurable via appsec.proxy.gke.gateway_classes. Multi-cluster GatewayClasses (names ending in -mc) are always skipped, including when added to that allowlist, because they require a net.gke.io ServiceImport callout backend that the Cluster Agent does not create. Each GCPTrafficExtension is owned by its Gateway via an owner reference, so Kubernetes garbage-collects it even if the Cluster Agent misses the Gateway deletion event.

Enhancement Notes

• Adds Cluster Agent telemetry for the DatadogInstrumentation controller, including the number of resources it tracks and reconciliation outcomes for checks and logs.
• Added the datadog.cluster_agent.autoscaling.workload.objective.target gauge, which exposes the target value configured in a DatadogPodAutoscaler spec.objectives. The metric is tagged with objective_type (pod_resource, container_resource or custom_query), value_type (utilization or absolute_value), objective_index (the 0-based position in spec.objectives that keeps each objective a distinct timeseries), and, for resource objectives, resource_name and kube_container_name.
• Collect NVIDIA Dynamo custom resources by default.
• Collect KubeRay RayCluster, RayCronJob, RayJob, and RayService custom resources by default.
• Single Step Instrumentation now evaluates local targeting (Helm, Operator, or datadog.yaml) before Remote Config policies. Local targets keep first-match-wins order. Remote Config policies use last-match-wins: the last matching policy applies, so a catch-all can be listed first and exceptions after. A workload that matches a local target is not overridden by a remote deny.
• Single Step Instrumentation now decides SSI versus local library injection from whether a configuration target or remote-config policy matched the pod, instead of a namespace-level eligibility approximation. Library annotations still short-circuit target selection for library versions and tracer configs (existing GA precedence).

Security Notes

• The Cluster Agent's admission controller webhook now enforces a size limit on incoming request bodies and validates the request content type before reading the body.
• The Cluster Agent no longer exposes the Go pprof profiling and expvar debug endpoints on its metrics port (metrics_port, default 5000) to remote callers. These /debug/ endpoints were previously served on all network interfaces without authentication; they are now restricted to loopback callers, and requests originating from any other address receive a 404. The /metrics endpoint is unchanged and remains reachable off-host so the node Agent can continue to scrape Cluster Agent telemetry. Local tooling such as the Cluster Agent flare, which connects over loopback, is unaffected.

Bug Fixes

• Fixed an issue where the DatadogPodAutoscaler controller could silently drop a status update after an HTTP 409 (Conflict) caused by a stale resourceVersion read from the informer cache. The reconcile now requeues on such a conflict so a subsequent pass retries the update with a refreshed object.
• Ensure Kubernetes endpoint check annotations take precedence over DatadogInstrumentation configurations that target the same Service and integration. This prevents duplicate endpoint checks and restores the CR-backed check when the overriding annotation is removed.
• Fix an issue on AKS clusters where the Cluster Agent and the AKS admission enforcer would repeatedly overwrite each other's changes to the datadog-webhook MutatingWebhookConfiguration/ValidatingWebhookConfiguration objects, causing the object has been modified errors to be logged in a loop. This affected admission controller features whose webhook rule did not otherwise restrict which namespaces it applies to (for example the DatadogInstrumentation CRD validating webhook), even when admission_controller.add_aks_selectors (DD_ADMISSION_CONTROLLER_ADD_AKS_SELECTORS) was enabled.
• Fix a crash in the kubernetes_state_core check (Cluster Agent or Cluster Check Runner) that occurred when using a wildcard entry ("*") in kubernetes_namespace_annotations_as_tags or the equivalent kubernetes_resources_annotations_as_tags namespace configuration, on any namespace without annotations.
• Fixed an issue where the Cluster Agent could schedule Prometheus Scrape OpenMetrics checks against Kubernetes services even when the autodiscovery configuration included kubernetes_container_names. The Cluster Agent now skips service and endpoint Prometheus Scrape scheduling for configurations that set kubernetes_container_names; scraping discovered by node Agents remains as is.
• Fix a Cluster Agent API bug that could log spurious superfluous response.WriteHeader call warnings. The internal telemetry wrapper now correctly tracks the response status when a handler writes the response body before explicitly setting the status code.
• Restore continuous kubernetes_state.endpoint.address_available and kubernetes_state.endpoint.address_not_ready reporting (including 0 for the opposite ready state). After the kube-state-metrics v2.18 bump, each metric was only emitted for addresses in that state, so healthy endpoints no longer reported address_not_ready=0 and fully unready endpoints no longer reported address_available=0.