NVIDIA/dcgm-exporter 4.8.4

on GitHub

Tested component versions: DCGM 4.6.1, go-dcgm v1.4611.2, go-nvml v0.13.4-0,
Go 1.27.1, and NVIDIA distroless CC v4.1.4.

Runtime and performance

• Release DCGM Exporter 4.8.4 with DCGM 4.6.1 and go-dcgm v1.4611.2.
• Add configurable DCGM field-watch retention limits through CLI, environment, and YAML configuration.
• Add the version 2 YAML schema with plural collections while preserving version 1 compatibility.
• Add configurable limits and request coalescing for concurrent /metrics scrapes.
• Reduce allocations during metric gathering, transformation, and rendering.
• Support metric configurations containing more than 128 DCGM fields by batching reads and splitting oversized field groups.
• Add optional GPU bind/unbind detection that resets DCGM and NVML, then rebuilds metrics for the new GPU topology (#706, reported by @varunrsekar; #716 by @dharmjit).
• Honor LD_LIBRARY_PATH when locating DCGM libraries.

Metrics and Kubernetes

• Add opt-in Go runtime, process, and HTTP handler metrics through --enable-exporter-metrics, DCGM_EXPORTER_ENABLE_EXPORTER_METRICS, or YAML configuration.
• Add opt-in GPU fabric health fields: DCGM_FI_DEV_FABRIC_HEALTH_SUMMARY and DCGM_FI_DEV_FABRIC_HEALTH_MASK.
• Add opt-in configured and enforced power-limit fields: DCGM_FI_DEV_POWER_MGMT_LIMIT and DCGM_FI_DEV_ENFORCED_POWER_LIMIT (#719 by @Kumaraditya18027).
• Refresh the optional NVLink error-counter definitions.
• Remove retired PCIe throughput fields from the supplied metric configurations. Custom configurations that use them fail and point to DCGM_FI_PROF_PCIE_TX_BYTES and DCGM_FI_PROF_PCIE_RX_BYTES.
• Collect compute-instance-scoped DCGM fields from MIG compute instances and add the GPU_CI_ID label while preserving whole-GPU collection.
• Attribute DRA MIG metrics to pods when the MIG device is visible to DCGM at startup, and refresh allocation metadata when ResourceSlices change (#714, reported by @guptaNswati).
• Require an exporter restart after live Dynamic MIG creation or removal; changing the MIG layout while the exporter is running is not supported.
• Add the optional pod_uid label to DRA-attributed metrics (#738 by @git-jxj; fixes #718).
• Reuse the kubelet pod-resources gRPC connection to prevent memory growth (#705 by @AshwinRamaniPsg).
• Prefer NVML MIG profile names and make MIG profile lookup available outside Kubernetes (#713, reported by @rahulait).
• Deduplicate repeated job IDs within each HPC mapping file while preserving mappings across files.
• Warn when HPC job mapping entries do not match discovered MIG instances.

Helm, containers, and packages

• Include all supported metric preset CSV files in binary packages.
• Include third-party notices in archives, packages, and container images.
• Include the dynamic linker cache in the distroless image so DCGM libraries can be found at startup (#724 by @arielrahamim; fixes #612).
• Add Helm values for DaemonSet annotations (#723 by @ranyhb).
• Add Helm values for readiness and liveness probe timing and failure thresholds (#679 by @ChenZiHong-Gavin).
• Render debug dump compression explicitly, including when it is disabled.
• Document opt-in scheduling on nodes with the nvidia.com/gpu:NoSchedule taint (#733 by @KR-Ravindra).
• Move raw Kubernetes deployment examples under deployment/examples.
• Update the Helm chart and deployment examples for DCGM Exporter 4.8.4.
• Update Go to 1.27.1 and refresh Go dependencies, container base images, Kubernetes tools, and development tooling.

Testing and documentation

• Expand CPU-only performance benchmarks for metric rendering and scraping.
• Expand coverage for field-watch retention, oversized field groups, concurrent scrapes, DRA allocation, pre-existing MIG devices, and GPU bind/unbind lifecycle handling.
• Simplify the repository README and direct users to the official DCGM Exporter installation, configuration, metrics, and release documentation.

Validation: make validate-versions, make test-main, make test-static,
make check-fmt, and make lint.