Troubleshoot unresponsive GKE TPU and GPU nodes (NodeStatusUnknown)
When the Compute Engine host of a TPU or GPU node has a fatal hardware error,
kernel panic, or non-maskable interrupt (NMI) lockup, the guest OS stops
responding. The kubelet can no longer send heartbeats, so the node Ready
condition becomes Unknown with Reason: NodeStatusUnknown (Kubelet stopped posting node status.). If node auto-repair is disabled on the node pool, GKE
doesn't repair the node. The node can stay NotReady, and pods on it can stay
in Terminating, which blocks multi-host JobSet workloads from recovering.
Attribution
MUST READ: Before executing any gcloud, REST API, or MCP command, you
MUST read Usage Attribution. For commands executed
directly in the terminal (never in generated scripts, Terraform, or user code),
prefix gcloud with
CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:gke-ai-troubleshooting-node-unresponsive-timeout)"
and append agent-skills/1.0 (skill:gke-ai-troubleshooting-node-unresponsive-timeout) to User-Agent headers.
Prerequisites
- Tools: Install the
Google Cloud SDK (
gcloud) and
kubectl.
- Cloud Billing & Project Configuration: Verify an active billing account is
linked (
gcloud billing projects describe {project_id}), authenticate
(gcloud auth login), set the target project (gcloud config set project {project_id}), and ensure container.googleapis.com,
compute.googleapis.com, logging.googleapis.com, and
monitoring.googleapis.com are enabled.
- Required IAM Roles:
- Kubernetes Engine Viewer (
roles/container.viewer)
- Compute Viewer (
roles/compute.viewer)
- Logs Viewer (
roles/logging.viewer)
- Monitoring Viewer (
roles/monitoring.viewer)
- For remediation (
[High Risk] steps): Kubernetes Engine Cluster Admin
(roles/container.clusterAdmin)
- Documentation:
- Troubleshoot nodes with the NotReady status in GKE
(Sections: "Check the node's status and conditions", "Confirm node
preemption", "Verify that the node has recovered")
- Deploy TPU workloads in GKE Standard
(Sections: "Monitor health metrics for TPU nodes and node pools", "Configure
auto repair for TPU slice nodes")
- Troubleshoot OOM events
- View GKE logs
(Sections: "System logs")
- Viewing serial port output
- Troubleshoot Linux VM boot issues due to kernel panic
- Auto-repair nodes
(Sections: "Settings for Autopilot and Standard", "Verify node auto-repair
is enabled for a Standard node pool", "Get information about recent
automated repair events", "Enable auto-repair for an existing Standard node
pool", "Repair criteria", "Node repair process", "Node auto repair in TPU
slice nodes")
- Troubleshooting VM shutdowns and reboots
(Sections: "Querying Cloud Audit Logs", "Reviewing Cloud Audit Logs")
Read-only rule: Run read-only diagnostic commands only. Never drain,
delete, or re-create nodes, or run any other command that changes the cluster.
Give the user any fix to apply themselves.
When you recommend a fix, link the doc section that describes it.
Diagnostic workflow
Step 0: Collect context and set the investigation window [Low Risk]
Collect the target parameters. By default, query a 60-minute window `[T - 30m, T
{project_id}: Google Cloud project ID
{cluster_name}: GKE cluster name
{location}: Cluster region or zone
{nodepool_name}: Target TPU or GPU node pool name
{node_name}: Unresponsive GKE node name (and its Compute Engine {zone})
{issue_time}: Incident timestamp in RFC3339 UTC
{start_time}: {issue_time} - 30m
{end_time}: {issue_time} + 30m
Step 1: Verify the NodeStatusUnknown heartbeat timeout [Low Risk]
- Check Kubernetes node conditions: To inspect the node status and verify
whether the
Ready condition is Unknown with Reason: NodeStatusUnknown
(Kubelet stopped posting node status.), follow the instructions in the
section
Check the node's status and conditions.
- Query Cloud Logging (read-only LQL): Query
k8s_node and k8s_cluster
logs across [{start_time}, {end_time}] to confirm when the control plane
lost heartbeat contact with {node_name}:
(resource.type="k8s_node" OR resource.type="k8s_cluster")
resource.labels.cluster_name="{cluster_name}"
("{node_name}" AND ("NodeNotReady" OR "NodeStatusUnknown" OR "Kubelet stopped posting node status"))
timestamp >= "{start_time}" AND timestamp <= "{end_time}"
- Query Cloud Monitoring (read-only PromQL): Correlate the duration of the
Unknown state using the GKE system metric
kubernetes.io/node/status_condition (kubernetes_io:node_status_condition,
GKE 1.32.1-gke.1357001+) documented in
Monitor health metrics for TPU nodes and node pools,
filtered by condition="Ready" and status="Unknown":
kubernetes_io:node_status_condition{
monitored_resource="k8s_node",
cluster_name="{cluster_name}",
node_name="{node_name}",
condition="Ready",
status="Unknown"
}
- Decision logic:
- If the node
Ready condition is True and NodeStatusUnknown is absent,
rule out an unresponsive node timeout and pivot to workload-level
troubleshooting (for example,
Troubleshoot OOM events)
rather than repairing or draining the node.
- If
Ready is Unknown (NodeStatusUnknown), proceed to Step 2.
Step 2: Inspect serial port output for a kernel panic [Low Risk]
The guest OS can't send logs after a fatal kernel freeze, so check the serial
port output of the node's VM:
- Cloud Logging: If serial port logging is enabled (i.e. VM metadata
serial-port-logging-enable is set to true), GKE system logs in Cloud
Logging include the node's serial port output. See the section
System logs.
Use Cloud Logging when the VM is stopped or has already been replaced by
auto-repair, or when you need more than the most recent output.
- Running VM: Follow
Viewing serial port output
to retrieve the serial port 1 output (
gcloud compute instances get-serial-port-output with --port=1) for {node_name} in {zone}. This
method returns only the most recent 1 MB of output per port.
Consult
Troubleshoot Linux VM boot issues due to kernel panic
to identify documented kernel panic and hardware crash patterns (such as Fatal Machine check, hung_task: blocked tasks, or NMI: Not continuing) in the
serial port output.
Step 3: Check node auto-repair status [Low Risk]
Find out why GKE hasn't repaired the unresponsive node:
Step 4: Check Compute Engine system events [Low Risk]
To list the system events for {node_name} around {issue_time}, follow the
instructions in the section
Querying Cloud Audit Logs.
Compare the method field with the table in the "Reviewing Cloud Audit Logs"
section of the same document, and look for:
compute.instances.hostError: a hardware or software issue on the physical
host caused the VM to crash.
compute.instances.preempted: Compute Engine preempted a Spot VM or
preemptible VM. For preempted nodes, also see
Confirm node preemption.
compute.instances.automaticRestart: Compute Engine restarted the VM after a
hostError or terminateOnHostMaintenance event.
compute.instances.guestTerminate: the VM's operating system initiated the
shutdown.
Step 5: Resolution [High Risk]
Guardrails:
- Never force-delete stuck
Terminating pods on an unresponsive node.
Force deletion doesn't wait for the kubelet to confirm that the pod has
stopped, so a replacement pod can start while the old one is still running.
See
Force Delete StatefulSet Pods.
- Never delete GKE-managed Compute Engine VM instances directly (
gcloud compute instances delete). Instead, check how long the node has reported
NodeStatusUnknown (Step 1), check the serial console output (Step 2), and
rely on node auto-repair, as described in the following steps.
- Enable node auto-repair (Standard clusters only):
- Let GKE repair the node:
- Explain that GKE repairs a node that reports
NotReady or no status for
the documented time threshold by draining and re-creating it, and link the
"Repair criteria" and "Node repair process" sections of
Auto-repair nodes.
GKE waits one hour for the drain to complete. If the drain doesn't
complete, GKE shuts the node down and creates a new node. Tell the user to
expect this one-hour drain wait before GKE re-creates the node.
- For multi-host TPU slice node pools, note per
Node auto repair in TPU slice nodes
that the entire node pool is re-created.
- Verify node recovery:
- After the repair completes, follow the instructions in the section
Verify that the node has recovered
to verify that the repaired node returns to
Ready status, then confirm
that the JobSet pods are running again.
1---2name: gke-ai-troubleshooting-node-unresponsive-timeout3description: >-4 Diagnose and mitigate GKE TPU or GPU nodes stuck in NotReady /5 NodeStatusUnknown ("Kubelet stopped posting node status") due to host kernel6 panics, hardware lockups, or disabled node auto-repair. Use when nodes stop7 heartbeating beyond the node auto-repair threshold and pods remain stuck in8 Terminating. Don't use for healthy nodes, pod-only application crashes, or9 routine GKE upgrades.10metadata:11 source: https://skillmd.com/skills/google/gke-ai-troubleshooting-node-unresponsive-timeout12 version: "1.0.1"13 category: Containers14---1516# Troubleshoot unresponsive GKE TPU and GPU nodes (`NodeStatusUnknown`)1718When the Compute Engine host of a TPU or GPU node has a fatal hardware error,19kernel panic, or non-maskable interrupt (NMI) lockup, the guest OS stops20responding. The kubelet can no longer send heartbeats, so the node `Ready`21condition becomes `Unknown` with `Reason: NodeStatusUnknown` (`Kubelet stopped22posting node status.`). If node auto-repair is disabled on the node pool, GKE23doesn't repair the node. The node can stay `NotReady`, and pods on it can stay24in `Terminating`, which blocks multi-host `JobSet` workloads from recovering.2526## Attribution2728**MUST READ**: Before executing any `gcloud`, REST API, or MCP command, you29MUST read [Usage Attribution](references/attribution.md). For commands executed30directly in the terminal (never in generated scripts, Terraform, or user code),31prefix `gcloud` with32`CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:gke-ai-troubleshooting-node-unresponsive-timeout)"`33and append `agent-skills/1.0 (skill:gke-ai-troubleshooting-node-unresponsive-timeout)` to `User-Agent` headers.3435## Prerequisites3637- **Tools**: Install the38 [Google Cloud SDK](https://cloud.google.com/sdk/docs/install) (`gcloud`) and39 `kubectl`.40- **Cloud Billing & Project Configuration**: Verify an active billing account is41 linked (`gcloud billing projects describe {project_id}`), authenticate42 (`gcloud auth login`), set the target project (`gcloud config set project43 {project_id}`), and ensure `container.googleapis.com`,44 `compute.googleapis.com`, `logging.googleapis.com`, and45 `monitoring.googleapis.com` are enabled.46- **Required IAM Roles**:47 - Kubernetes Engine Viewer (`roles/container.viewer`)48 - Compute Viewer (`roles/compute.viewer`)49 - Logs Viewer (`roles/logging.viewer`)50 - Monitoring Viewer (`roles/monitoring.viewer`)51 - For remediation (`[High Risk]` steps): Kubernetes Engine Cluster Admin52 (`roles/container.clusterAdmin`)53- **Documentation**:54 - [Troubleshoot nodes with the NotReady status in GKE](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/node-notready.md.txt)55 (Sections: "Check the node's status and conditions", "Confirm node56 preemption", "Verify that the node has recovered")57 - [Deploy TPU workloads in GKE Standard](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt)58 (Sections: "Monitor health metrics for TPU nodes and node pools", "Configure59 auto repair for TPU slice nodes")60 - [Troubleshoot OOM events](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/oom-events.md.txt)61 - [View GKE logs](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/view-logs.md.txt)62 (Sections: "System logs")63 - [Viewing serial port output](https://docs.cloud.google.com/compute/docs/troubleshooting/viewing-serial-port-output.md.txt)64 - [Troubleshoot Linux VM boot issues due to kernel panic](https://docs.cloud.google.com/compute/docs/troubleshooting/kernel-panic.md.txt)65 - [Auto-repair nodes](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt)66 (Sections: "Settings for Autopilot and Standard", "Verify node auto-repair67 is enabled for a Standard node pool", "Get information about recent68 automated repair events", "Enable auto-repair for an existing Standard node69 pool", "Repair criteria", "Node repair process", "Node auto repair in TPU70 slice nodes")71 - [Troubleshooting VM shutdowns and reboots](https://docs.cloud.google.com/compute/docs/troubleshooting/troubleshooting-reboots.md.txt)72 (Sections: "Querying Cloud Audit Logs", "Reviewing Cloud Audit Logs")7374> **Read-only rule**: Run read-only diagnostic commands only. Never drain,75> delete, or re-create nodes, or run any other command that changes the cluster.76> Give the user any fix to apply themselves.7778When you recommend a fix, link the doc section that describes it.7980---8182## Diagnostic workflow8384### Step 0: Collect context and set the investigation window `[Low Risk]`8586Collect the target parameters. By default, query a 60-minute window `[T - 30m, T87+ 30m]` around `{issue_time}`:8889- `{project_id}`: Google Cloud project ID90- `{cluster_name}`: GKE cluster name91- `{location}`: Cluster region or zone92- `{nodepool_name}`: Target TPU or GPU node pool name93- `{node_name}`: Unresponsive GKE node name (and its Compute Engine `{zone}`)94- `{issue_time}`: Incident timestamp in RFC3339 UTC95- `{start_time}`: `{issue_time} - 30m`96- `{end_time}`: `{issue_time} + 30m`9798---99100### Step 1: Verify the `NodeStatusUnknown` heartbeat timeout `[Low Risk]`1011021. **Check Kubernetes node conditions**: To inspect the node status and verify103 whether the `Ready` condition is `Unknown` with `Reason: NodeStatusUnknown`104 (`Kubelet stopped posting node status.`), follow the instructions in the105 section106 [Check the node's status and conditions](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/node-notready.md.txt).1072. **Query Cloud Logging (read-only LQL)**: Query `k8s_node` and `k8s_cluster`108 logs across `[{start_time}, {end_time}]` to confirm when the control plane109 lost heartbeat contact with `{node_name}`:110111```112(resource.type="k8s_node" OR resource.type="k8s_cluster")113resource.labels.cluster_name="{cluster_name}"114("{node_name}" AND ("NodeNotReady" OR "NodeStatusUnknown" OR "Kubelet stopped posting node status"))115timestamp >= "{start_time}" AND timestamp <= "{end_time}"116```1171183. **Query Cloud Monitoring (read-only PromQL)**: Correlate the duration of the119 `Unknown` state using the GKE system metric120 `kubernetes.io/node/status_condition` (`kubernetes_io:node_status_condition`,121 GKE `1.32.1-gke.1357001+`) documented in122 [Monitor health metrics for TPU nodes and node pools](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt),123 filtered by `condition="Ready"` and `status="Unknown"`:124125```promql126kubernetes_io:node_status_condition{127 monitored_resource="k8s_node",128 cluster_name="{cluster_name}",129 node_name="{node_name}",130 condition="Ready",131 status="Unknown"132}133```134135- **Decision logic**:136 - If the node `Ready` condition is `True` and `NodeStatusUnknown` is absent,137 **rule out** an unresponsive node timeout and pivot to workload-level138 troubleshooting (for example,139 [Troubleshoot OOM events](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/oom-events.md.txt))140 rather than repairing or draining the node.141 - If `Ready` is `Unknown` (`NodeStatusUnknown`), proceed to Step 2.142143---144145### Step 2: Inspect serial port output for a kernel panic `[Low Risk]`146147The guest OS can't send logs after a fatal kernel freeze, so check the serial148port output of the node's VM:149150- **Cloud Logging**: If serial port logging is enabled (i.e. VM metadata151 `serial-port-logging-enable` is set to `true`), GKE system logs in Cloud152 Logging include the node's serial port output. See the section153 [System logs](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/view-logs.md.txt).154 Use Cloud Logging when the VM is stopped or has already been replaced by155 auto-repair, or when you need more than the most recent output.156- **Running VM**: Follow157 [Viewing serial port output](https://docs.cloud.google.com/compute/docs/troubleshooting/viewing-serial-port-output.md.txt)158 to retrieve the serial port 1 output (`gcloud compute instances159 get-serial-port-output` with `--port=1`) for `{node_name}` in `{zone}`. This160 method returns only the most recent 1 MB of output per port.161162Consult163[Troubleshoot Linux VM boot issues due to kernel panic](https://docs.cloud.google.com/compute/docs/troubleshooting/kernel-panic.md.txt)164to identify documented kernel panic and hardware crash patterns (such as `Fatal165Machine check`, `hung_task: blocked tasks`, or `NMI: Not continuing`) in the166serial port output.167168---169170### Step 3: Check node auto-repair status `[Low Risk]`171172Find out why GKE hasn't repaired the unresponsive node:173174- **Autopilot clusters**: Autopilot always repairs nodes, and you can't turn175 this off (see176 [Settings for Autopilot and Standard](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt)).177 Skip the configuration check and check the repair history.178- **Standard clusters**: Follow179 [Verify node auto-repair is enabled for a Standard node pool](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt)180 to check whether `autoRepair` is enabled on `{nodepool_name}`.181- **Repair history (both modes)**: Auto-repair can be enabled and the repair can182 still fail. Per183 [Configure auto repair for TPU slice nodes](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt),184 check the repair status, including the failure reason, in the operation185 history described in186 [Get information about recent automated repair events](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt).187 If the failure is caused by insufficient quota, tell the user to contact their188 Google Cloud account representative to increase the quota.189190---191192### Step 4: Check Compute Engine system events `[Low Risk]`193194To list the system events for `{node_name}` around `{issue_time}`, follow the195instructions in the section196[Querying Cloud Audit Logs](https://docs.cloud.google.com/compute/docs/troubleshooting/troubleshooting-reboots.md.txt).197Compare the `method` field with the table in the "Reviewing Cloud Audit Logs"198section of the same document, and look for:199200- `compute.instances.hostError`: a hardware or software issue on the physical201 host caused the VM to crash.202- `compute.instances.preempted`: Compute Engine preempted a Spot VM or203 preemptible VM. For preempted nodes, also see204 [Confirm node preemption](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/node-notready.md.txt).205- `compute.instances.automaticRestart`: Compute Engine restarted the VM after a206 `hostError` or `terminateOnHostMaintenance` event.207- `compute.instances.guestTerminate`: the VM's operating system initiated the208 shutdown.209210---211212### Step 5: Resolution `[High Risk]`213214> **Guardrails**:215> - **Never** force-delete stuck `Terminating` pods on an unresponsive node.216> Force deletion doesn't wait for the kubelet to confirm that the pod has217> stopped, so a replacement pod can start while the old one is still running.218> See219> [Force Delete StatefulSet Pods](https://kubernetes.io/docs/tasks/run-application/force-delete-stateful-set-pod/#force-deletion).220> - **Never** delete GKE-managed Compute Engine VM instances directly (`gcloud221> compute instances delete`). Instead, check how long the node has reported222> `NodeStatusUnknown` (Step 1), check the serial console output (Step 2), and223> rely on node auto-repair, as described in the following steps.2242251. **Enable node auto-repair (Standard clusters only)**:226 - Autopilot clusters always auto-repair nodes, so skip this step.227 - Before the user enables auto-repair on a multi-host TPU slice node pool,228 tell them that GKE re-creates the entire node pool when a node in it needs229 repair (see step 2).230 - If `autoRepair` is disabled on `{nodepool_name}`, link the user to231 [Enable auto-repair for an existing Standard node pool](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt)232 and233 [Configure auto repair for TPU slice nodes](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt)234 so they can apply the change themselves.2352. **Let GKE repair the node**:236 - Explain that GKE repairs a node that reports `NotReady` or no status for237 the documented time threshold by draining and re-creating it, and link the238 "Repair criteria" and "Node repair process" sections of239 [Auto-repair nodes](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt).240 GKE waits one hour for the drain to complete. If the drain doesn't241 complete, GKE shuts the node down and creates a new node. Tell the user to242 expect this one-hour drain wait before GKE re-creates the node.243 - For multi-host TPU slice node pools, note per244 [Node auto repair in TPU slice nodes](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/node-auto-repair.md.txt)245 that the entire node pool is re-created.2463. **Verify node recovery**:247 - After the repair completes, follow the instructions in the section248 [Verify that the node has recovered](https://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/node-notready.md.txt)249 to verify that the repaired node returns to `Ready` status, then confirm250 that the `JobSet` pods are running again.