Uninstall

How to uninstall Nsight Operator.

Uninstalling Nsight Operator

helm uninstall nsight-operator -n nsight-operator

helm uninstall removes the Deployments, Services, ConfigMaps, RBAC, and the default CRs created by the chart.

Note

In multi-tenant mode, tenant namespaces may hold auto-provisioned or manually deployed CRs (NsightCoordinator, NsightGateway, etc.) that are independent of the Helm release. See Quickstart Example (Multitenant Mode) for tenant-side cleanup.

Reclaiming storage

helm uninstall does not delete PersistentVolumeClaims created by StatefulSets (MinIO when persistence is enabled, the OTel collector). Delete them manually to reclaim the underlying storage.

Warning

Deleting these PVCs is irreversible and permanently destroys their contents – including MinIO-hosted .nsys-rep profiling reports and any buffered OTel collector data. Before running the command below, confirm the reports are no longer needed, or back up the data to external S3 / a different cluster. There is no helm rollback for PVC contents once they are deleted.

kubectl delete pvc -l app.kubernetes.io/name=nsight-operator -n nsight-operator

Removing the Dynamo helper configuration

If you used nsight_operator_dynamo.py to install (single-tenant) or configure (multi-tenant) a namespace, run the matching teardown subcommand in addition to helm uninstall so that Dynamo-managed resources – the NsightOperatorProfileConfig CR, the GPU-metrics DaemonSet and its ConfigMap, and any tenant-namespace Secrets / ConfigMaps the script created – are removed:

# Single-tenant
python3 nsight_operator_dynamo.py -n nsight-operator uninstall

# Multi-tenant (per tenant namespace)
python3 nsight_operator_dynamo.py -n <tenant-ns> unconfigure

If you only used the Helm chart (no Dynamo helper), helm uninstall on its own is sufficient.

Air-gapped clusters

helm uninstall schedules a pre-delete Helm Job that runs a small distroless Python image to remove dynamically created CRs before the chart’s RBAC is torn down. By default that image is pulled from nvcr.io/nvidia/distroless/python; in air-gapped clusters with no egress to nvcr.io the Job’s Pod stays in ImagePullBackOff and helm uninstall hangs until it times out.

To avoid this, mirror the distroless Python image to your private registry and override global.nsightCloud.distrolessPythonImage and global.nsightCloud.distrolessPythonImagePullSecrets if the mirror requires authentication at install time, so that the value is already baked into the rendered pre-delete Job when uninstall runs:

helm install nsight-operator <chart> -n nsight-operator \
  --set global.nsightCloud.distrolessPythonImage=registry.example.internal/nvidia/distroless/python:3.14-v4.0.9 \
  --set global.nsightCloud.distrolessPythonImagePullSecrets[0].name=my-mirror-pull-secret