Troubleshooting

Get help with common questions, errors, and warnings

  18 minute read  

Common troubleshooting steps

Installation

Workloads and recommendations

Errors and warnings

For status details, see Workload recommendation statuses.

Common troubleshooting steps

Most troubleshooting involves the checks listed below.

Agent troubleshooting

View high-level information such as Agent status and cluster connection status

Check the Agent pod status:

kubectl get pods -n stormforge-system -l app.kubernetes.io/component=cluster-agent

View the agent pod logs:

kubectl logs -n stormforge-system -l app.kubernetes.io/component=cluster-agent

Use the optional CLI image to check the cluster connection status:

kubectl run stormforge-check-connect --attach --rm --restart=Never --quiet \
  -n stormforge-system --image registry.stormforge.io/library/stormforge-cli \
  --override-type=strategic \
  --overrides='{"spec": {
                  "serviceAccount": "stormforge-cluster-agent",
                  "containers": [{
                    "name": "stormforge-check-connect",
                    "envFrom": [
                      {"secretRef": {"name": "stormforge-auth"}},
                      {"secretRef": {"name": "stormforge-env"}}]}]}}' \
   -- check connect

Examine the output to determine:

  • Whether the agent pod is running
  • Whether the agent pod is reporting any errors
  • Whether the application can connect to the api.stormforge.io endpoint

Check whether the Agent is running

kubectl get pods -n stormforge-system -l app.kubernetes.io/component=cluster-agent

View the Agent logs

These logs show output and errors related to the Agent Deployments.

  • If the Agent won’t start (for example, connectivity issues when using a proxy), check the init container status:

    kubectl logs -n stormforge-system -l app.kubernetes.io/component=cluster-agent -c init --tail=-1

  • If the Agent is running:

    kubectl logs -n stormforge-system -l app.kubernetes.io/component=cluster-agent --tail=-1

Describe the Agent Pod and events

kubectl describe pod -n stormforge-system -l app.kubernetes.io/component=cluster-agent

Examine the output:

  • Check the pod status, state and reason, conditions, and events.
  • In the .containers spec, check the arguments, state, liveness, and readiness.

View the ConfigMaps settings

kubectl get configmaps -n stormforge-system -l app.kubernetes.io/name=stormforge -o yaml

Scan the relevant ConfigMaps for any unusual or unexpected settings. To make changes, identify the correct Helm values needed to change the ConfigMap and set them by running helm upgrade. If you make changes by editing the ConfigMaps directly, those manual edits will be overwritten on upgrade.

Restart the Agent

In some scenarios, restarting the Agent solves the problem. Remember that this will result in a few moments of downtime while new pods scale up:

kubectl rollout restart -n stormforge-system deployment.apps/stormforge-cluster-agent

To check the status of the restart, run:

kubectl rollout status deployment -n stormforge-system stormforge-cluster-agent -w

A successful restart generates this message:
deployment "stormforge-cluster-agent" successfully rolled out

Applier troubleshooting

View the Applier logs

These logs contain output and errors related to the Applier Deployment.

kubectl logs -n stormforge-system -l component=applier --tail=-1

Describe the Applier Pod and events

kubectl describe pod -n stormforge-system -l component=applier

Examine the output:

  • Check the pod status, state and reason, conditions, and events.
  • In the .containers spec, check the arguments, state, liveness, and readiness.

Node Agent troubleshooting

The Node Agent is optional and disabled by default. These steps apply only if you enabled it to collect GPU metrics.

Check whether the Node Agent is running

kubectl get pods -n stormforge-system -l app.kubernetes.io/component=node-agent -o wide

Expect one pod per GPU node. The -o wide output shows which node each pod is on.

View the Node Agent logs

kubectl logs -n stormforge-system -l app.kubernetes.io/component=node-agent --tail=-1

Describe the Node Agent Pod and events

kubectl describe pod -n stormforge-system -l app.kubernetes.io/component=node-agent

Examine the output:

  • Check the pod status, state and reason, conditions, and events.
  • In the .containers spec, check the arguments, state, liveness, and readiness.

Restart the Node Agent

kubectl rollout restart daemonset/stormforge-node-agent -n stormforge-system

No Node Agent pods on the GPU nodes

If the DaemonSet reports 0 desired pods, or pods are missing from nodes that have GPUs, the node selector or the tolerations don’t match those nodes.

  1. Check what the DaemonSet expects:

    kubectl get daemonset stormforge-node-agent -n stormforge-system \
  -o jsonpath='{.spec.template.spec.nodeSelector}{"\n"}{.spec.template.spec.tolerations}{"\n"}'

  2. Check the labels and taints on a GPU node, replacing NODE_NAME with the node name:

    kubectl get node NODE_NAME -o jsonpath='{.metadata.labels}{"\n"}{.spec.taints}{"\n"}'

  3. If they don’t match, set nodeAgent.nodeSelector and nodeAgent.tolerations to match your GPU nodes. See Node selection.

Node Agent is running but GPU data is missing

If Node Agent pods are running on your GPU nodes but GPU workloads show no GPU data:

  1. Confirm that StormForge identifies the workload as a GPU workload:

    stormforge view workloads --gpu

    If the workload isn’t listed, see How GPU workloads are identified.

  2. Check the Node Agent logs for GPU initialization warnings:

    kubectl logs -n stormforge-system -l app.kubernetes.io/component=node-agent --tail=-1 --prefix \
  | grep -E "gpu provider init failed|dcgm profiling unavailable"

    • gpu provider init failed: the Node Agent couldn’t initialize the NVIDIA driver library on that node, so it isn’t collecting GPU metrics. If the line includes ERROR_LIBRARY_NOT_FOUND, confirm that the node runs the NVIDIA container runtime and that your cluster policies admit the Node Agent’s SYS_ADMIN capability. See Permissions and host access. If the node has no GPU, adjust Node selection so that the Node Agent doesn’t run there.

    • dcgm profiling unavailable: engine activity metrics, such as tensor core activity, are missing. Memory, power, and utilization metrics are unaffected. See How GPU metrics are collected.

  3. Check the Node Agent’s collection health metrics. Forward its metrics port to your machine, replacing POD_NAME with a Node Agent pod name:

    kubectl port-forward -n stormforge-system pod/POD_NAME 9401:9401

    In a second terminal, query the metrics:

    curl -s localhost:9401/metrics \
  | grep -E "^sf_gpu_(scrape_errors|allocation_scrape_errors|allocated_unresolved_devices)"

    If you changed nodeAgent.metricsPort, use that port instead of 9401. If the command prints nothing and the port-forward shows no errors, the Node Agent isn’t collecting GPU metrics on that node. Check its logs as described in the previous step.

    Metric A value above 0 means What to do
    sf_gpu_scrape_errors The Node Agent’s last read from the NVIDIA driver failed. Check the Node Agent logs for gpu scrape.
    sf_gpu_allocation_scrape_errors The Node Agent’s last read of GPU allocations from the kubelet failed, so the GPU tab is missing data. Check the Node Agent logs for pod-resources scrape. If the value stays above 0, confirm that the kubelet serves the PodResources API socket under /var/lib/kubelet/pod-resources on the node.
    sf_gpu_allocated_unresolved_devices Some allocated GPUs couldn’t be matched to a physical GPU, so they’re missing from the affected workloads’ GPU tab. The reason label is dra_slice for a GPU allocated through Dynamic Resource Allocation, or mig_geometry for a MIG device. For dra_slice, a value can appear briefly after the Node Agent starts. If it stays above 0, confirm that the cluster runs Kubernetes 1.34 or later. For mig_geometry, restart the Node Agent. If either value stays above 0, contact StormForge support.

Installation

Health check failed with a Failed to create StormForge client message

This message can occur when installing either the Agent or Applier, and it indicates a connectivity problem between your environment and the StormForge environment. For example, a proxy or a firewall that isn’t configured correctly can cause timeout, proxy, and TLS issues.

Start by checking the connectivity to the StormForge endpoints:

kubectl run stormforge-check-connect --attach --rm --restart=Never --quiet \
  -n stormforge-system --image registry.stormforge.io/library/stormforge-cli \
  --override-type=strategic \
  --overrides='{"spec": {
                  "serviceAccount": "stormforge-cluster-agent",
                  "containers": [{
                    "name": "stormforge-check-connect",
                    "envFrom": [
                      {"secretRef": {"name": "stormforge-auth"}},
                      {"secretRef": {"name": "stormforge-env"}}]}]}}' \
    -- check connect

If any endpoint status in the output is not OK, try the following steps:

  1. Run the following busybox and wget commands to check for a network issue such as a problematic network policy, firewall, or egress controller:

    kubectl -n stormforge-system run -i --tty busybox --image=busybox --restart=Never -- sh

    At the busybox command prompt, enter:

    wget https://api.stormforge.io

    In this example, although it is counterintuitive, the 404 error indicates connectivity to StormForge servers:

    % kubectl -n stormforge-system run -i --tty busybox --image=busybox --restart=Never -- sh
If you don't see a command prompt, try pressing enter.
/ # 
/ # wget https://api.stormforge.io
Connecting to api.stormforge.io (XX.XXX.XXX.XXX:443)
wget: note: TLS certificate validation not implemented
wget: server returned error: HTTP/1.1 404 Not Found

    Type exit to exit busybox and then delete the pod:

    kubectl delete pod busybox -n stormforge-system

  2. If the busybox and wget commands above do not return a 404 error, run the following command, which uses a netshoot image to determine which network hop is dropping the connection.
    Note: You must be connected to the cluster with admin rights and the pod must run in privileged mode (.spec.privileged: true):

    kubectl -n stormforge-system run tmp-shell --rm -i --tty --image nicolaka/netshoot

    At the [tmp-shell] prompt, enter:

    tcptraceroute api.stormforge.io 443

    The following example illustrates end-to-end connectivity. If you are troubleshooting, you might see * * * indefinitely at a particular hop.

    % kubectl -n stormforge-system run tmp-shell --rm -i --tty --image nicolaka/netshoot
tmp-shell  ~  tcptraceroute api.stormforge.io 443
Selected device dev0, address 192.168.0.3, port 33305 for outgoing packets
Tracing the path to api.stormforge.io (198.51.100.1) on TCP port 443 (HTTPS), 30 hops max
1  192.168.0.1  0.056 ms  0.052 ms  0.048 ms
2  192.168.0.2  0.274 ms  0.215 ms  0.205 ms
3  * * *
4  ec2-198-51-100-1.compute-1.amazonaws.com (198.51.100.1)  0.787 ms  0.678 ms  0.542 ms
5  ec2-198-51-100-1.compute-1.amazonaws.com (198.51.100.1)  1.520 ms  0.832 ms  0.817 ms
6  ec2-198-51-100-1.compute-1.amazonaws.com (198.51.100.1) [open]  1.288 ms  2.021 ms  2.333 ms

The installation failed. What can I check?

Agent

If the Agent didn’t install correctly, you might not be able to:

Here are a few things you can check:

Applier

The Applier uses the secret provided by the Agent — check for the secret’s existence by running:

kubectl get deployment,secret -n stormforge-system

Check the output for the following item:

NAME                           TYPE     DATA   AGE
secret/stormforge-auth         Opaque   5      31d

If the secret is missing, reinstall StormForge using helm install stormforge oci://registry.stormforge.io/library/stormforge.

Workloads and recommendations

I don’t see the workloads (or namespaces) that I expect on the dashboard. What can I check?

Reasons why you might not see some workloads or namespaces include:

  • The namespace exists but it is empty - no pods, no built-in types.
  • The namespace contains only batch jobs (which are not supported) or pods without an owner reference (this is rare).
  • The namespace has no workloads in a healthy state: for example, a DaemonSet isn’t fully operational (which occurs when not all are healthy).
  • The namespace allowNamespaces or denyNamespaces Helm chart values constrain which namespaces to optimize. No namespaces are excluded by default, so this applies only if one of these lists has been set. Check with helm get values.

There are a few things you can check: pods on the cluster, the Agent logs, and the allow/deny namespaces lists.

Get a list of all running pods on the cluster

kubectl get all -A

You might choose to modify this command or add other arguments, based on your familiarity with Kubernetes.

In the output:

  • In the NAME column, look at the pod list (pod/POD_NAME rows) and make sure the stormforge-cluster-agent-POD_SUFFIX pod and your application pods in other namespaces are running.
  • In the NAME column, look for batch.apps: If it exists, make sure that the corresponding namespace contains other Kubernetes types (deployment.apps, replicaset.apps, daemonset.apps).
  • For each pod whose STATUS column that has a status of RUNNING, check the READY column: Make sure all the nodes are ready. For example, 7/7 means the workload is healthy (all 7 pods are running); 3/7 is unhealthy.

Check whether there are any workloads on the cluster
Run kubectl get deployments (include the -n NAMESPACE argument if you’re not checking the default namespace). If there are no deployments listed in the output, StormForge won’t collect any metrics.

  • If you need a sample workload, you can download the StormForge sample application, which installs one workload on your cluster. From a command line, run: 
    helm install optlive-showcase-app oci://registry.stormforge.io/examples/optlive-showcase-app
    Learn more about the sample application in this guide.

Check the Agent logs

  1. Get the Agent logs as described above.
  2. Scan the output for errors and any unusual or unexpected settings.

Check whether the allowList or denyList constraints are too restrictive

  1. Get the StormForge ConfigMap as described above.
  2. Find cluster-agent.yml in the output, and look for the allowNamespaces or denyNamespaces argument.
    • If either list is present and doesn’t match what you expect, adjust it as needed. If both are present, the allowList has precedence and the denyList is ignored.
    • If neither list is included in this spec, then neither list has been set, and something else is causing the problem.

How do I grant RBAC permissions on a patch target?

To apply or export the recommended workload settings, StormForge must:

  • Support the patch target type.
  • Have RBAC access permissions on the patch target.

When these permissions are missing, you’ll see a message similar to this:

please grant rbac access permissions on name/namespace: %s/%s, kind: %s, apiversion: %s

To grant RBAC access permissions:

  1. Find the values.yaml file from the Helm chart that was used to install StormForge.

    • If you don’t have the file:

      1. Download the chart:

        helm show values oci://registry.stormforge.io/library/stormforge

        If you use a proxy server or private container registry, follow the steps in Advanced installation.

      2. Copy the clusterAgent.rbac section into a new file named rbac-update.yaml. Proceed to the next step.

  2. In the clusterAgent.rbac.additionalRules section, add a .apiGroups section that includes the resources listed in the message.

    • Remember to replace API_VERSION and KIND with the values from the please grant rbac access permissions ... message.

      clusterAgent:
  rbac:
    additionalRules: |
      - apiGroups:
        - API_VERSION
        resources:
        - KIND
        verbs:
        - get
        - list
        - watch

  3. Apply the changes by running one of the following commands:

    • If you edited the values.yaml file, run:

      helm upgrade --values.yaml

    • If you created a separate file named rbac-update.yaml, run:

      helm upgrade --values.yaml -f rbac-update.yaml

I didn’t get a recommendation for a workload. What can I check?

Check the basics:

  • Look for more information in a message in the StormForge banner.
  • Check that the recommendation schedule is set to the frequency that you expect:

    1. Click Workloads in left navigation and drill down to the workload.

    2. Click Settings > Schedule and check the Recommendation Schedule.

      To change the schedule, set the live.stormforge.io/schedule annotation as described in the Schedule topic.

Check for scale-down events: You won’t get a recommendation if the workload has been scaled to zero replicas for the entire 7-day period.

Recommendations don’t seem right based on the CPU/memory utilization. What can I do?

StormForge responds to changes in a workload’s profile (CPU/memory utilization) and generates recommendations based on those changes and recent data.

If a planned change to a workload dramatically alters the profile, StormForge Support can help you restart the workload’s learning period (initial observation period). StormForge will collect fresh metrics and generate new preliminary recommendations, which you can apply on demand, for the duration of the learning period.

When the learning period is complete, StormForge generates recommendations that can be applied automatically if auto-deploy is enabled, or you can apply them on demand.

Related information:

My cluster is disconnected. What can I check?

If StormForge doesn’t get a response from a cluster after about one hour, it marks the cluster as disconnected (). The data shown is the last view that StormForge has of the data.

If you believe the cluster should be connected, check the Agent status and cluster connection status. Restart the Agent if needed.

Errors and warnings

Failed to send batch, retrying

Post "https://in.stormforge.io/prometheus/write": oauth2: "too_many_attempts" We have detected suspicious login behavior and further attempts will be blocked. Please contact the administrator.

If enough incorrect client_ID/client_secret combinations originate from one IP address, any further attempts to connect to StormForge are blocked.

Troubleshooting steps:

  1. Determine which authentication setup you’re using in your estate. An administrator typically makes this decision, usually before installing StormForge on an initial cluster.

    • One clientID/clientSecret pair used by all clusters
    • One clientID/clientSecret pair per cluster

  2. On each cluster, locate your access credentials file (the .yaml file that contains your clientID and clientSecret). This file was created when you either:

    • Added a cluster by clicking Add Cluster in the StormForge UI
    • Ran stormforge auth create CREDENTIAL_NAME > CREDENTIAL_FILE from the command line before you installed the StormForge Agent

Errors and warnings are grouped by type when you hover over the status icon in a workload list.

Recommender failure

Failed to generate recommendation (1000)

This error can occur for different reasons. Typically, contacting support is the best action to take, but you can also try one of the following:

Scaled to zero

No recommendation is generated when a workload scales to zero for too long.

This workload has been scaled to zero for the entire last 7 days. No non-zero replica data is available to generate a recommendation.

If the scaling down is unexpected, investigate the workload using kubectl commands.

Missing data

No recommendation was generated because data is missing or is stale. When this happens, you’ll see one of the following errors:

Expired

The recommendation was not computed in a timely manner.
You can either wait for the next recommendation to be generated or generate a new one by: - Clicking Regenerate on the workload details page and then applying the recommendation.
- Running stormforge create recommendations [flags]. Append the --autodeploy flag to apply it immediately, or run stormforge apply --recommendation RECOMMENDATION_NAME [flags] to apply it manually.

Applier failure

If a workload doesn’t restart or function as expected after a recommendation is applied, the StormForge Applier rolls back the workload to its last known good state and logs one of the following messages in the Applier logs.

  • applied patch but timed out waiting for workload to become ready
    rolled back to previous state not waiting for workload to become ready
    workload was unhealthy prior to patch

    The workload wasn’t healthy even before applying any changes, and you should investigate the workload directly.

    Start by describing the pods relating to the workload resource and examine the Events section for clues.

    kubectl describe pod NAME

    You can learn more about troubleshooting common issues with Kubernetes resources in the Kubernetes docs.

  • applied patch but timed out waiting for workload to become ready
    rolled back to previous state workload returned to ready

    or

    applied patch but timed out waiting for workload to become ready rolled back to previous state timed out waiting for pods to be ready

    While rollback is sometimes a desirable safety feature to have enabled, at scale and in large environments these errors often occur when the Applier’s monitoring period of 5min isn’t long enough for workloads to become Ready. The Applier determines that the change is a failure and rolls the workload back to the previous state.

    If this error happens often, disable the rollback feature.

    Disable rollback

    1. Add the following to your values.yaml: 
      additionalEnvVars:
- name: STORMFORGE_SKIP_ROLLBACK
  value: "true"
    2. Upgrade StormForge: 
      helm upgrade stormforge oci://registry.stormforge.io/library/stormforge \
  -n stormforge-system \
  --reset-then-reuse-values \
  -f values.yaml

Missing patch path for a Job

StormForge marks a recommendation for a standalone Job as not patchable with the reason MissingPatchPath and a message such as:

  • The workload is missing podtemplatemetadata patchPath configuration for its Job patchTarget.
  • The workload is missing CPU requests patchPath configuration for its Job patchTarget.

Kubernetes doesn’t allow changes to a Job’s Pod template after creation, so StormForge can’t patch the Job. Use a webhook-based apply method to set recommended values on Pods at admission:

  1. Enable the Applier webhook.
  2. Add the following annotation to the Job: 
    live.stormforge.io/apply.method: "WebhookHybridRollout"

If the Job uses WebhookImmediateRollout, setting live.stormforge.io/pod-template.metadata.patch-path: "-" on the Job has the same effect. See Pod template metadata.

Unsupported workloads

Workloads whose status is Unsupported () are configured in such a way that StormForge does not generate recommendations for them. The following messages shown for unsupported workloads are informational only, and typically, no action is required.

Job doesn’t run long enough to make a recommendation

A CronJob (or any short-duration Job) must typically run for at least 10 continuous minutes in order for StormForge to generate a recommendation. In some scenarios, 5-10 minutes is sufficient to generate a recommendation, depending on the frequency of the Job.

If the Job does not provide enough metrics within 5-10 minutes, StormForge labels the workload as Unsupported.

StormForge does not generate recommendations for Jobs that run for less than 5 minutes.

Corrupt data detected for container

No recommendation was generated because some of the metrics data StormForge has collected exhibit signs of corruption that make the metrics data unusable, such as metrics showing impossibly high CPU or memory usage. If you see workloads with this Unsupported reason, please contact Support.

Oscillating replicas

No recommendation was generated because the HPA oscillates between min and max replicas. If this is unexpected, investigate the workload. Increasing the max replicas can sometimes eliminate the oscillation, enabling StormForge to generate recommendations.

Target utilization max

No recommendation was generated because the HPA target.averageUtilization is set to greater than or equal to 100.

Last modified October 5, 2026