Observability & Cost

Autopilot surfaces cluster and managed-service health, network traffic, and cost attribution in the same dashboard — all from existing Prometheus metrics and the GCP Cloud Monitoring API, with no extra agents.

GCP managed services monitoring

Autopilot provides read-only monitoring for GCP services associated with each cluster. The controller polls the GCP Cloud Monitoring API every 60s and surfaces health and metrics in the dashboard.

Service	Metrics	Health thresholds
Cloud SQL	CPU, disk, connections, state	CPU >90%, disk >80/95%, connections >80/95%
Memorystore Redis	Memory, evicted keys, hit ratio, state	Memory >80/95%, evictions >0, hit ratio `<50%`
Cloud NAT	Port usage, dropped packets, NAT IPs	Ports >80/95%, drops ≥1/s, IPs >40
Cloud Storage	Size, request rates, error rates	Error rate >1%, high request rates
Cloud DNS	Query rates, latency, DNSSEC status	Latency thresholds, DNSSEC validation failures
VPC Peering	Peering status, route counts	Inactive peerings, route quota proximity
Artifact Registry	Image count, vulnerabilities, size	Critical/high vulnerabilities, scan failures
Secret Manager	Secret count, rotation status, expiry	Expired secrets, missing rotation

Configuration

Per-cluster in deploy/envs/<cluster>.yaml:

gcpServices:
  enabled: true
  services:
    - type: CloudSQL
      instanceName: my-database
    - type: MemorystoreRedis
      instanceName: my-cache
    - type: CloudNAT
      routerName: my-router

The GCP Services page shows a card for each service with health indicators and key metrics. Click through for a detail page with usage charts and events.

Network monitoring

Cluster-wide network observability powered by Prometheus. The Network page provides real-time visibility into traffic patterns, errors, and drops across the cluster without any additional agents or configuration — it queries existing Prometheus container network metrics. What it shows:

Cluster throughput — aggregate ingress/egress bandwidth (bytes/sec) with a time-series chart.
Top namespaces by traffic — which namespaces generate the most network I/O.
Top pods by traffic — individual pods ranked by bandwidth consumption.
Per-node network health — errors/sec and drops/sec per node, flagging nodes with problems.
Error summary — cluster-wide error and drop rates with node counts.

Key design details:

Virtual interfaces (veth, calico, flannel, CNI bridges) are excluded — only physical NIC traffic is counted, to avoid double-counting.
Data is cached with singleflight so dashboard polls don’t hammer Prometheus.
No configuration needed — it works automatically if Prometheus is available.

Navigate to Infrastructure → Network. The page shows summary cards, a throughput time-series chart, and tables for top namespaces, top pods, and node health.

Package cost tracking

Per-package cost attribution for tenant namespaces. Each pod’s share of node cost is computed using request-weighted blended CPU/memory:

pod_cost = node_hourly_cost × (0.5 × pod_cpu_req/node_allocatable_cpu
                              + 0.5 × pod_mem_req/node_allocatable_mem)

Pods are grouped by the xano.com/package label. Unattributed cost is tracked as spillover in three categories: DaemonSet pods, system pods, and idle capacity. Cost snapshots are recorded every 6 hours for trend analysis. The Package Costs page (Infrastructure section) shows a cost breakdown table, spillover analysis, and a historical cost chart.

System health checks

The health endpoint (GET /api/v1/health) runs a battery of checks:

Check	What it validates
Kubernetes API	API server reachable, response time
Prometheus	Reachable, query latency, sample availability
CRDs	All CRD types registered, at least one policy exists
Metrics Server	K8s metrics-server responding (optional)
Node Capacity	Managed nodes within scaling policy limits
Prometheus Storage	PVC utilization below thresholds
Resizer Effectiveness	PodResizePolicy producing recommendations
Scaler Freshness	NodeScalingPolicy evaluating on schedule
Metrics Quality	Node count in Prometheus matches the K8s API
OOM Death Spirals	No pods in repeated OOM-restart loops
CoreDNS	CoreDNS pods healthy in kube-system
Webhook	Mutation webhook registered and serving
Nodes Not Ready	All nodes in Ready condition
Ingress Controller	Ingress controller pods healthy
GCP Services	Managed service health
CAST AI Conflict	No conflicting CAST AI agents detected

Get Started

Platform

Operations

Migrating from CAST AI

GCP managed services monitoring

Configuration

Network monitoring

Package cost tracking

System health checks

​GCP managed services monitoring

​Configuration

​Network monitoring

​Package cost tracking

​System health checks

GCP managed services monitoring

Configuration

Network monitoring

Package cost tracking

System health checks