GKE JobSet Interruption Troubleshooting
Use this skill to systematically diagnose and resolve JobSet interruptions,
restarts, and preemptions on GKE clusters hosting large-scale AI/ML workloads.
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-jobset-interruption)"
and append agent-skills/1.0 (skill:gke-ai-troubleshooting-jobset-interruption) to User-Agent headers.
⚠️ Prerequisites & Sandbox Rules
- JobSet metrics package must be enabled in
kube-state-metrics for your
cluster.
- Cloud Logging and Cloud Monitoring enabled for the Google Cloud Project.
- Autonomous & Sandbox Execution Rule: If API queries, tools, or commands
encounter
403 Permission Denied, authentication errors, or network
isolation, do NOT enter authentication or credential troubleshooting
loops. Populate the query templates with the acquired variables
({project_id}, {cluster_name}, {workload_name}, {start_time},
{end_time}), inspect any locally staged telemetry or mock data files if
available, and complete the diagnostic workflow and resolution
recommendations autonomously.
🔍 Diagnostic Workflow
Step 0: Context Acquisition & Time Window Definition
Independently gather context using tools, workspace files, environment details,
or user prompt context:
- Project ID (
{project_id})
- Cluster Name (
{cluster_name})
- Workload Name (JobSet Name) (
{workload_name})
- Workload Namespace (
{namespace})
- Issue Time (
{issue_time})
If specific variables are not explicitly provided by the user, inspect cluster
resources or logs to determine them, or use the {variable} placeholders
provided.
Time Handling Rules
- Autonomous Time Window: If a relative time (e.g., "X minutes ago") or no
exact timestamp is provided, calculate the query window based on current
time or available log timestamps.
- Window Calculation: If a timestamp
{issue_time} is available (or
calculated as T), set {start_time} = T - 30m and {end_time} = T + 30m.
Step 1: Identify JobSet Restarts and Attempts [Low Risk]
Verify if the JobSet is experiencing restart loops and determine the frequency
of restarts.
Visual Chart / MQL Query - restarts
PromQL Metric Query - restarts
PromQL Query Specification:
kube_jobset_restarts{jobset_name="{workload_name}", cluster="{cluster_name}"}
Diagnostic Logic: A non-zero or increasing value for restarts indicates
that the JobSet is being actively restarted by the controller due to worker
failure or interruption.
Automation: Proceed to Step 2 automatically after reporting findings.
Step 2: Inspect Nodepool Interruptions [Low Risk]
Determine if the JobSet restarts were triggered by physical nodepool-level
events (such as spot preemptions, maintenance, or host terminations).
A. Metrics Query (Nodepool Interruption Counts)
Visual Chart / MQL Query - interruptions
MQL Query Specification:
fetch k8s_node_pool
| metric 'kubernetes.io/node_pool/interruption_count'
| filter cluster_name == '{cluster_name}'
| align next_older(10m)
| every 10m
| group_by [metric.interruption_type, metric.interruption_reason, metadata.system.node_pool_name], [val: sum(value)]
PromQL Query - interruptions
PromQL Query Specification:
sum by (interruption_type, interruption_reason, node_pool_name, cluster_name) (
avg_over_time(kubernetes_io:node_pool_interruption_count{cluster_name="{cluster_name}"}[10m])
)
B. Log Query (Nodepool Life Events)
LQL Log Filter Specification:
resource.type="gke_nodepool"
AND resource.labels.cluster_name="{cluster_name}"
AND timestamp >= "{start_time}"
AND timestamp <= "{end_time}"
Diagnostic Logic:
- PreemptionEvent: Spot VMs were preempted, or node was scale-down.
- MaintenanceEvent: Node pool updated or Google scheduled maintenance.
- TerminationEvent: Serious host failures. Check
interruption_reason
or logs for host issues.
- See Failure Signatures for examples
of node termination logs and preemption events.
Automation: Proceed to Step 3 automatically.
Step 3: Inspect Nodes and Underlying Host VMs [Low Risk]
Correlate node readiness failures with physical host VMs to see if a single
faulty host repeatedly fails coordinator pods.
A. Metrics Query (Node Ready Status Check)
Visual Chart / MQL Query - node status
MQL Query Specification:
fetch k8s_node
| metric 'kubernetes.io/node/status_condition'
| filter cluster_name == '{cluster_name}' && metric.condition == 'Ready' && metric.status == 'False'
| align next_older(1m)
| every 1m
| group_by [node_name, metadata.user.gke_nodepool], [val: max(value)]
PromQL Query - node status
PromQL Query Specification:
sum by (status, condition, node_pool_name) (
kubernetes_io:node_status_condition{cluster_name="{cluster_name}", condition="Ready", status="False"}
)
B. Metrics Query (Node-to-Host Metadata Topology Correlation)
MQL Query Specification:
fetch k8s_node
| metric 'kubernetes.io/node/cpu/total_cores'
| filter cluster_name == '{cluster_name}'
| align next_older(1m)
| every 1m
| group_by [node_name, metadata.user.gce_topology_host, metadata.user.gke_nodepool], [val: max(value)]
C. Log Query (Node Fault Logs)
LQL Log Filter Specification:
resource.type="k8s_node"
AND resource.labels.cluster_name="{cluster_name}"
AND (textPayload:"host error" OR textPayload:"kernel panic" OR textPayload:"hardware failure" OR textPayload:"NodeNotReady")
AND timestamp >= "{start_time}"
AND timestamp <= "{end_time}"
Diagnostic Logic: Identify if specific nodes are unhealthy
(Ready=False or Unknown) and correlate them to their GCE physical host
ID via metadata.user.gce_topology_host. Check if the same host is
repeatedly failing.
Automation: Proceed to Step 4 automatically.
Step 4: Inspect Pod and Worker / Container Failures [Low Risk]
Analyze pod status phases and retrieve coordinator worker logs to identify
application-level crashes or network deadlocks.
Required Execution Order: You MUST analyze pod status phases (Section A)
and unschedulable pod metrics (Section B) to assess overall workload health
before inspecting specific worker container logs (Section C).
A. Metrics Query (Pod Lifecycle Phases)
Visual Chart / MQL Query - pod phase
PromQL Query - pod phase
PromQL Query Specification:
sum by (phase) (
avg_over_time(kube_pod_status_phase{cluster="{cluster_name}", pod=~"{workload_name}.*"}[10m])
)
B. Metrics Query (Unschedulable Pod Count)
C. Log Query (Worker Container Logs)
LQL Log Filter Specification:
resource.type="k8s_container"
AND resource.labels.cluster_name="{cluster_name}"
AND labels."k8s-pod/jobset_sigs_k8s_io/jobset-name"="{workload_name}"
AND timestamp >= "{start_time}"
AND timestamp <= "{end_time}"
Diagnostic Logic:
- Check the pod timeline to spot pending or unschedulable pods.
- Use worker container logs to analyze worker 0 in slice 0 (coordinator)
for NCCL timeouts, collective communication issues, or MegaScale hangs.
Automation: Proceed to Resolution.
🛠️ Resolution Workflow
Resolution 1: Preemption & Autoscaling Optimizations [Low Risk]
If Step 2 showed high preemption counts on Spot VMs:
- Action: Suggest switching critical long-running training workloads to
GKE Reserved/On-Demand VMs or utilizing Compact Placement Policies
to minimize defragmentation interruptions.
- Justification: Eliminates spot-market preemptions and reduces training
restarts.
Resolution 2: Quarantine Faulty Host VMs [High Risk]
If Step 3 identified a specific host ID (gce-topology-host) that consistently
fails or triggers restarts across multiple attempts:
- Action: Recommend cordoning/draining the GKE node, deleting the
underlying GCE VM instance to trigger instance recreation, and opening a
support ticket with Google Cloud Support specifying the physical host ID.
- Justification: GKE auto-repair will recreate the VM instance on healthy
physical hardware, preventing infinite restart loops.
📋 Copypaste Checklist
1---2name: gke-ai-troubleshooting-jobset-interruption3metadata:4 source: https://skillmd.com/skills/google/gke-ai-troubleshooting-jobset-interruption5 version: "1.0.1"6 category: Containers7description: >-8 Diagnoses GKE JobSet interruptions, restarts, and preemptions for AI/ML training workloads autonomously.9 Use when troubleshooting JobSet restart loops, spot VM preemptions, node readiness failures, host VM issues, or coordinator worker crashes.10 Don't use for general GKE cluster creation, basic workload deployment, or non-JobSet application issues.11---1213# GKE JobSet Interruption Troubleshooting1415Use this skill to systematically diagnose and resolve JobSet interruptions,16restarts, and preemptions on GKE clusters hosting large-scale AI/ML workloads.1718## Attribution1920**MUST READ**: Before executing any `gcloud`, REST API, or MCP command, you21MUST read [Usage Attribution](references/attribution.md). For commands executed22directly in the terminal (never in generated scripts, Terraform, or user code),23prefix `gcloud` with24`CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:gke-ai-troubleshooting-jobset-interruption)"`25and append `agent-skills/1.0 (skill:gke-ai-troubleshooting-jobset-interruption)` to `User-Agent` headers.2627## ⚠️ Prerequisites & Sandbox Rules2829- JobSet metrics package must be enabled in `kube-state-metrics` for your30 cluster.31- Cloud Logging and Cloud Monitoring enabled for the Google Cloud Project.32- **Autonomous & Sandbox Execution Rule**: If API queries, tools, or commands33 encounter `403 Permission Denied`, authentication errors, or network34 isolation, **do NOT enter authentication or credential troubleshooting35 loops**. Populate the query templates with the acquired variables36 (`{project_id}`, `{cluster_name}`, `{workload_name}`, `{start_time}`,37 `{end_time}`), inspect any locally staged telemetry or mock data files if38 available, and complete the diagnostic workflow and resolution39 recommendations autonomously.4041## 🔍 Diagnostic Workflow4243### Step 0: Context Acquisition & Time Window Definition4445Independently gather context using tools, workspace files, environment details,46or user prompt context:4748- **Project ID** (`{project_id}`)49- **Cluster Name** (`{cluster_name}`)50- **Workload Name (JobSet Name)** (`{workload_name}`)51- **Workload Namespace** (`{namespace}`)52- **Issue Time** (`{issue_time}`)5354If specific variables are not explicitly provided by the user, inspect cluster55resources or logs to determine them, or use the `{variable}` placeholders56provided.5758#### Time Handling Rules59601. **Autonomous Time Window**: If a relative time (e.g., "X minutes ago") or no61 exact timestamp is provided, calculate the query window based on current62 time or available log timestamps.632. **Window Calculation**: If a timestamp `{issue_time}` is available (or64 calculated as `T`), set `{start_time}` = `T - 30m` and `{end_time}` = `T +65 30m`.6667--------------------------------------------------------------------------------6869### Step 1: Identify JobSet Restarts and Attempts [Low Risk]7071Verify if the JobSet is experiencing restart loops and determine the frequency72of restarts.7374#### Visual Chart / MQL Query - restarts7576- **MQL Query Specification**:7778 ```mql79 fetch prometheus_target80 | metric 'prometheus.googleapis.com/kube_jobset_restarts/gauge'81 | filter resource.cluster_name == '{cluster_name}' && metric.jobset_name == '{workload_name}'82 | align next_older(1m)83 | every 1m84 | group_by [metric.jobset_name], [val: max(value)]85 ```8687#### PromQL Metric Query - restarts8889- **PromQL Query Specification**:9091 ```promql92 kube_jobset_restarts{jobset_name="{workload_name}", cluster="{cluster_name}"}93 ```9495- **Diagnostic Logic**: A non-zero or increasing value for restarts indicates96 that the JobSet is being actively restarted by the controller due to worker97 failure or interruption.9899- **Automation**: Proceed to Step 2 automatically after reporting findings.100101--------------------------------------------------------------------------------102103### Step 2: Inspect Nodepool Interruptions [Low Risk]104105Determine if the JobSet restarts were triggered by physical nodepool-level106events (such as spot preemptions, maintenance, or host terminations).107108#### A. Metrics Query (Nodepool Interruption Counts)109110##### Visual Chart / MQL Query - interruptions111112- **MQL Query Specification**:113114 ```mql115 fetch k8s_node_pool116 | metric 'kubernetes.io/node_pool/interruption_count'117 | filter cluster_name == '{cluster_name}'118 | align next_older(10m)119 | every 10m120 | group_by [metric.interruption_type, metric.interruption_reason, metadata.system.node_pool_name], [val: sum(value)]121 ```122123##### PromQL Query - interruptions124125- **PromQL Query Specification**:126127 ```promql128 sum by (interruption_type, interruption_reason, node_pool_name, cluster_name) (129 avg_over_time(kubernetes_io:node_pool_interruption_count{cluster_name="{cluster_name}"}[10m])130 )131 ```132133#### B. Log Query (Nodepool Life Events)134135- **LQL Log Filter Specification**:136137 ```sql138 resource.type="gke_nodepool"139 AND resource.labels.cluster_name="{cluster_name}"140 AND timestamp >= "{start_time}"141 AND timestamp <= "{end_time}"142 ```143144- **Diagnostic Logic**:145146 - **PreemptionEvent**: Spot VMs were preempted, or node was scale-down.147 - **MaintenanceEvent**: Node pool updated or Google scheduled maintenance.148 - **TerminationEvent**: Serious host failures. Check `interruption_reason`149 or logs for host issues.150 - See [Failure Signatures](references/failure_signatures.md) for examples151 of node termination logs and preemption events.152153- **Automation**: Proceed to Step 3 automatically.154155--------------------------------------------------------------------------------156157### Step 3: Inspect Nodes and Underlying Host VMs [Low Risk]158159Correlate node readiness failures with physical host VMs to see if a single160faulty host repeatedly fails coordinator pods.161162#### A. Metrics Query (Node Ready Status Check)163164##### Visual Chart / MQL Query - node status165166- **MQL Query Specification**:167168 ```mql169 fetch k8s_node170 | metric 'kubernetes.io/node/status_condition'171 | filter cluster_name == '{cluster_name}' && metric.condition == 'Ready' && metric.status == 'False'172 | align next_older(1m)173 | every 1m174 | group_by [node_name, metadata.user.gke_nodepool], [val: max(value)]175 ```176177##### PromQL Query - node status178179- **PromQL Query Specification**:180181 ```promql182 sum by (status, condition, node_pool_name) (183 kubernetes_io:node_status_condition{cluster_name="{cluster_name}", condition="Ready", status="False"}184 )185 ```186187#### B. Metrics Query (Node-to-Host Metadata Topology Correlation)188189- **MQL Query Specification**:190191 ```mql192 fetch k8s_node193 | metric 'kubernetes.io/node/cpu/total_cores'194 | filter cluster_name == '{cluster_name}'195 | align next_older(1m)196 | every 1m197 | group_by [node_name, metadata.user.gce_topology_host, metadata.user.gke_nodepool], [val: max(value)]198 ```199200#### C. Log Query (Node Fault Logs)201202- **LQL Log Filter Specification**:203204 ```sql205 resource.type="k8s_node"206 AND resource.labels.cluster_name="{cluster_name}"207 AND (textPayload:"host error" OR textPayload:"kernel panic" OR textPayload:"hardware failure" OR textPayload:"NodeNotReady")208 AND timestamp >= "{start_time}"209 AND timestamp <= "{end_time}"210 ```211212- **Diagnostic Logic**: Identify if specific nodes are unhealthy213 (`Ready=False` or `Unknown`) and correlate them to their GCE physical host214 ID via `metadata.user.gce_topology_host`. Check if the same host is215 repeatedly failing.216217- **Automation**: Proceed to Step 4 automatically.218219--------------------------------------------------------------------------------220221### Step 4: Inspect Pod and Worker / Container Failures [Low Risk]222223Analyze pod status phases and retrieve coordinator worker logs to identify224application-level crashes or network deadlocks.225226> **Required Execution Order**: You MUST analyze pod status phases (Section A)227> and unschedulable pod metrics (Section B) to assess overall workload health228> before inspecting specific worker container logs (Section C).229230#### A. Metrics Query (Pod Lifecycle Phases)231232##### Visual Chart / MQL Query - pod phase233234- **MQL Query Specification**:235236 ```mql237 fetch k8s_pod238 | metric 'kubernetes.io/pod/status/phase'239 | filter cluster_name == '{cluster_name}' && pod_name ==~ '{workload_name}.*'240 | align next_older(10m)241 | every 10m242 | group_by [metric.phase], [val: count()]243 ```244245##### PromQL Query - pod phase246247- **PromQL Query Specification**:248249 ```promql250 sum by (phase) (251 avg_over_time(kube_pod_status_phase{cluster="{cluster_name}", pod=~"{workload_name}.*"}[10m])252 )253 ```254255#### B. Metrics Query (Unschedulable Pod Count)256257- **MQL Query Specification**:258259 ```mql260 fetch k8s_pod261 | metric 'kubernetes.io/pod/status/unschedulable'262 | filter cluster_name == '{cluster_name}' && pod_name ==~ '{workload_name}.*'263 | align next_older(10m)264 | every 10m265 | group_by [pod_name], [val: max(value)]266 ```267268#### C. Log Query (Worker Container Logs)269270- **LQL Log Filter Specification**:271272 ```sql273 resource.type="k8s_container"274 AND resource.labels.cluster_name="{cluster_name}"275 AND labels."k8s-pod/jobset_sigs_k8s_io/jobset-name"="{workload_name}"276 AND timestamp >= "{start_time}"277 AND timestamp <= "{end_time}"278 ```279280- **Diagnostic Logic**:281282 1. Check the pod timeline to spot pending or unschedulable pods.283 2. Use worker container logs to analyze worker 0 in slice 0 (coordinator)284 for NCCL timeouts, collective communication issues, or MegaScale hangs.285286- **Automation**: Proceed to Resolution.287288--------------------------------------------------------------------------------289290## 🛠️ Resolution Workflow291292### Resolution 1: Preemption & Autoscaling Optimizations [Low Risk]293294If Step 2 showed high preemption counts on Spot VMs:295296- **Action**: Suggest switching critical long-running training workloads to297 **GKE Reserved/On-Demand VMs** or utilizing **Compact Placement Policies**298 to minimize defragmentation interruptions.299- **Justification**: Eliminates spot-market preemptions and reduces training300 restarts.301302### Resolution 2: Quarantine Faulty Host VMs [High Risk]303304If Step 3 identified a specific host ID (`gce-topology-host`) that consistently305fails or triggers restarts across multiple attempts:306307- **Action**: Recommend cordoning/draining the GKE node, deleting the308 underlying GCE VM instance to trigger instance recreation, and opening a309 support ticket with Google Cloud Support specifying the physical host ID.310- **Justification**: GKE auto-repair will recreate the VM instance on healthy311 physical hardware, preventing infinite restart loops.312313--------------------------------------------------------------------------------314315## 📋 Copypaste Checklist316317- [ ] Gather context and compute `{start_time}` (`{issue_time} - 30m`) and318 `{end_time}` (`{issue_time} + 30m`) window.319- [ ] Query JobSet restart attempts.320- [ ] Check Nodepool interruptions (spot preemptions vs. hardware321 terminations).322- [ ] Query node-to-host mapping and check node logs for physical host errors.323- [ ] Inspect pod timeline status and coordinator worker container logs.324- [ ] Recommend appropriate scheduling strategy (On-demand vs Spot) or host VM325 quarantining.