/gke-ai-troubleshooting-tpu-metrics-monitoring

name: gke-ai-troubleshooting-tpu-metrics-monitoring
description: >-
  Monitors and troubleshoots GKE TPU workloads, nodes, and node pools using GKE
  system metrics and PromQL. Use when monitoring TensorCore duty cycle, TPU memory,
  node readiness, multi-host TPU node pool availability, host maintenance or
  preemption interruptions, and calculating MTTR or MTBI metrics for GKE TPUs.
  Don't use for general non-TPU GKE workload monitoring or non-metric TPU debugging.
metadata:
  category: CloudObservabilityAndMonitoring

GKE TPU Metrics Monitoring Guide

This skill enables the agent to monitor GKE TPU workloads, nodes, and node pools using GKE system metrics. It helps diagnose if workload interruptions or performance issues are caused by underlying infrastructure.

Step 0: Mandatory Context

Independently gather required context (such as cluster details or node pool names) using available GKE and Cloud tools, or use the provided `{variable}` placeholders:

• `{project_id}`: The GCP Project ID.
• `{cluster_name}`: The GKE Cluster Name.
• `{location}`: The GKE Cluster Location (region or zone).
• `{node_name}`: (Optional) The name of the specific GKE node.
• `{node_pool_name}`: (Optional) The name of the GKE node pool.

---

Diagnostic Steps

Step 1: Verify TPU Runtime Metrics Configuration [Low Risk] [Auto]

Before analyzing runtime metrics, verify that the workload is configured to export them. This ensures the cluster and container environment are set up for automated metric scraping and visibility into accelerator health.

• **Action**: Verify that the Pod specification and cluster meet the following prerequisites:
• `containerPort: 8431` exposed on the TPU container (required for Prometheus metric scraping).
• JAX version `0.4.14` or later if using JAX (earlier versions do not export runtime metrics).
• GKE version is `1.27.4-gke.900` or later (required for TPU runtime metric support).
• GKE System Metrics are enabled on the cluster (required for Cloud Monitoring ingestion).

Step 2: Monitor TPU Runtime Metrics [Low Risk] [Auto]

If configured correctly, the following metrics are available in Cloud Monitoring (monitored resources `k8s_node` and `k8s_container`):

• **Container Metrics**:
• `kubernetes.io/container/accelerator/duty_cycle`: Percentage of time over the past sampling period (60 seconds) during which the TensorCores were actively processing on a TPU chip.
• `kubernetes.io/container/accelerator/memory_used`: Amount of accelerator memory allocated in bytes.
• `kubernetes.io/container/accelerator/memory_total`: Total accelerator memory in bytes.
• **Node Metrics**:
• `kubernetes.io/node/accelerator/duty_cycle`
• `kubernetes.io/node/accelerator/memory_used`
• `kubernetes.io/node/accelerator/memory_total`

Step 3: Check Node Status Condition [Low Risk] [Auto]

Query the status condition of GKE nodes (GKE version `1.32.1-gke.1357001` or later).

• **PromQL Query (Check if a specific node is Ready)**:

  kubernetes_io:node_status_condition{monitored_resource="k8s_node", cluster_name="{cluster_name}", node_name="{node_name}", condition="Ready", status="True"}

• **PromQL Query (List nodes with non-Ready conditions that are True)**:

  kubernetes_io:node_status_condition{monitored_resource="k8s_node", cluster_name="{cluster_name}", condition!="Ready", status="True"}

• **PromQL Query (List nodes that are NOT Ready)**:

  kubernetes_io:node_status_condition{monitored_resource="k8s_node", cluster_name="{cluster_name}", condition="Ready", status="False"}

• **PromQL Query (Fleet-wide node status)**:

  avg by (condition,status)(avg_over_time(kubernetes_io:node_status_condition{monitored_resource="k8s_node"}[5m]))

Step 4: Check Node Pool Status [Low Risk] [Auto]

Query the status of multi-host TPU node pools.

• **PromQL Query (Verify if a specific node pool is Running)**:

  kubernetes_io:node_pool_status{monitored_resource="k8s_node_pool", cluster_name="{cluster_name}", node_pool_name="{node_pool_name}", status="Running"}

• **PromQL Query (Monitor node pools grouped by status)**:

  count by (status)(count_over_time(kubernetes_io:node_pool_status{monitored_resource="k8s_node_pool"}[5m]))

_Possible statuses_: `Provisioning`, `Running`, `Error`, `Reconciling`, `Stopping`.

Step 5: Check Node Pool Availability [Low Risk] [Auto]

Query if all nodes in a multi-host TPU node pool are available.

• **PromQL Query (Check availability over time)**:

  avg by (node_pool_name)(avg_over_time(kubernetes_io:node_pool_multi_host_available{monitored_resource="k8s_node_pool", cluster_name="{cluster_name}"}[5m]))

_Value_: `1` (True, all nodes available) or `0` (False, some nodes unavailable).

Step 6: Analyze Node Interruptions [Low Risk] [Auto]

Query the count of interruptions for GKE nodes.

• **PromQL Query (Breakdown of interruptions and causes)**:

  sum by (interruption_type,interruption_reason)(sum_over_time(kubernetes_io:node_interruption_count{monitored_resource="k8s_node"}[5m]))

_Interruption Types_: `TerminationEvent`, `MaintenanceEvent`, `PreemptionEvent`. _Interruption Reasons_: `HostError`, `Eviction`, `AutoRepair`.

• **PromQL Query (Filter for Host Maintenance events)**:

  sum by (interruption_type,interruption_reason)(sum_over_time(kubernetes_io:node_interruption_count{monitored_resource="k8s_node", interruption_reason="HW/SW Maintenance"}[5m]))

• **PromQL Query (Interruption count aggregated by node pool)**:

  sum by (node_pool_name,interruption_type,interruption_reason)(sum_over_time(kubernetes_io:node_pool_interruption_count{monitored_resource="k8s_node_pool", interruption_reason="HW/SW Maintenance", node_pool_name="{node_pool_name}"}[5m]))

Step 7: Calculate Recovery and Interruption Metrics [Low Risk] [Auto]

Calculate Mean Time to Recovery