Metrics Setup

Configure Prometheus, ServiceMonitor, and custom metrics for autoscaling

Overview

The metrics setup enables autoscaling by collecting Lightning ASR metrics with Prometheus and exposing them to Kubernetes HPA through the Prometheus Adapter.

Architecture

/* The original had a syntax error in Mermaid—edges must connect nodes, not labels. “Metrics” is now a node, and edge directions/names are consistent. */

Components

Prometheus

Collects and stores metrics from Lightning ASR pods.

Included in chart:

values.yaml

scaling:
  auto:
    enabled: true

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      serviceMonitorSelectorNilUsesHelmValues: false
      retention: 7d
      resources:
        requests:
          memory: 2Gi

ServiceMonitor

CRD that tells Prometheus which services to scrape.

Enabled for Lightning ASR:

values.yaml

scaling:
  auto:
    lightningAsr:
      servicemonitor:
        enabled: true

Prometheus Adapter

Converts Prometheus metrics to Kubernetes custom metrics API.

Configuration:

values.yaml

prometheus-adapter:
  prometheus:
    url: http://smallest-prometheus-stack-prometheus.default.svc
    port: 9090
  rules:
    custom:
      - seriesQuery: "asr_active_requests"
        resources:
          overrides:
            namespace: {resource: "namespace"}
            pod: {resource: "pod"}
        name:
          matches: "^(.*)$"
          as: "${1}"
        metricsQuery: "asr_active_requests{<<.LabelMatchers>>}"

Available Metrics

Lightning ASR exposes the following metrics:

Metric	Type	Description
asr_active_requests	Gauge	Current number of active transcription requests
asr_total_requests	Counter	Total requests processed
asr_failed_requests	Counter	Total failed requests
asr_request_duration_seconds	Histogram	Request processing time
asr_model_load_time_seconds	Gauge	Time to load model on startup
asr_gpu_utilization	Gauge	GPU utilization percentage
asr_gpu_memory_used_bytes	Gauge	GPU memory used

Verify Metrics Setup

Check Prometheus

Forward Prometheus port:

kubectl port-forward -n default svc/smallest-prometheus-stack-prometheus 9090:9090

Open http://localhost:9090 and verify:

1. Status → Targets: Lightning ASR endpoints should be “UP”
2. Graph: Query asr_active_requests - should return data
3. Status → Service Discovery: Should show ServiceMonitor

Check ServiceMonitor

kubectl get servicemonitor -n smallest

Expected output:

NAME            AGE
lightning-asr   5m

Describe ServiceMonitor:

kubectl describe servicemonitor lightning-asr -n smallest

Should show:

Spec:
  Endpoints:
    Port: metrics
    Path: /metrics
  Selector:
    Match Labels:
      app: lightning-asr

Check Prometheus Adapter

Verify custom metrics are available:

kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1" | jq -r '.resources[].name' | grep asr

Expected output:

pods/asr_active_requests
pods/asr_total_requests
pods/asr_failed_requests

Query specific metric:

kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1/namespaces/smallest/pods/*/asr_active_requests" | jq .

Custom Metric Configuration

Add New Custom Metrics

To expose additional metrics to HPA:

values.yaml

prometheus-adapter:
  rules:
    custom:
      - seriesQuery: "asr_active_requests"
        resources:
          overrides:
            namespace: {resource: "namespace"}
            pod: {resource: "pod"}
        name:
          matches: "^(.*)$"
          as: "${1}"
        metricsQuery: "asr_active_requests{<<.LabelMatchers>>}"
      
      - seriesQuery: "asr_gpu_utilization"
        resources:
          overrides:
            namespace: {resource: "namespace"}
            pod: {resource: "pod"}
        name:
          as: "gpu_utilization"
        metricsQuery: "avg_over_time(asr_gpu_utilization{<<.LabelMatchers>>}[2m])"

External Metrics

For cluster-wide metrics:

values.yaml

prometheus-adapter:
  rules:
    external:
      - seriesQuery: 'kube_deployment_status_replicas{deployment="lightning-asr"}'
        metricsQuery: 'sum(kube_deployment_status_replicas{deployment="lightning-asr"})'
        name:
          as: "lightning_asr_replica_count"
        resources:
          overrides:
            namespace: {resource: "namespace"}

Use in HPA:

spec:
  metrics:
    - type: External
      external:
        metric:
          name: lightning_asr_replica_count
        target:
          type: Value
          value: "5"

Prometheus Configuration

Retention Policy

Configure how long metrics are stored:

values.yaml

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      retention: 15d
      retentionSize: "50GB"

Storage

Persist Prometheus data:

values.yaml

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      storageSpec:
        volumeClaimTemplate:
          spec:
            storageClassName: gp3
            accessModes: ["ReadWriteOnce"]
            resources:
              requests:
                storage: 100Gi

Scrape Interval

Adjust how frequently metrics are collected:

values.yaml

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      scrapeInterval: 30s
      evaluationInterval: 30s

Lower intervals (e.g., 15s) provide faster HPA response but increase storage.

Recording Rules

Pre-compute expensive queries:

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      additionalScrapeConfigs:
        - job_name: 'lightning-asr-aggregated'
          scrape_interval: 15s
          static_configs:
            - targets: ['lightning-asr:2269']
      
      additionalPrometheusRulesMap:
        asr-rules:
          groups:
            - name: asr_aggregations
              interval: 30s
              rules:
                - record: asr:requests:rate5m
                  expr: rate(asr_total_requests[5m])
                
                - record: asr:requests:active_avg
                  expr: avg(asr_active_requests) by (namespace)
                
                - record: asr:gpu:utilization_avg
                  expr: avg(asr_gpu_utilization) by (namespace)

Use recording rules in HPA for better performance.

Alerting Rules

Create alerts for anomalies:

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      additionalPrometheusRulesMap:
        asr-alerts:
          groups:
            - name: asr_alerts
              rules:
                - alert: HighErrorRate
                  expr: rate(asr_failed_requests[5m]) > 0.1
                  for: 5m
                  labels:
                    severity: warning
                  annotations:
                    summary: "High ASR error rate"
                    description: "Error rate is {{ $value }} errors/sec"
                
                - alert: HighQueueLength
                  expr: asr_active_requests > 50
                  for: 2m
                  labels:
                    severity: warning
                  annotations:
                    summary: "ASR queue backing up"
                    description: "{{ $value }} requests queued"
                
                - alert: GPUMemoryHigh
                  expr: asr_gpu_memory_used_bytes / 24000000000 > 0.9
                  for: 5m
                  labels:
                    severity: warning
                  annotations:
                    summary: "GPU memory usage high"
                    description: "GPU memory at {{ $value | humanizePercentage }}"

Debugging Metrics

Check Metrics Endpoint

Directly query Lightning ASR metrics:

kubectl port-forward -n smallest svc/lightning-asr 2269:2269
curl http://localhost:2269/metrics

Expected output:

# HELP asr_active_requests Current active requests
# TYPE asr_active_requests gauge
asr_active_requests{pod="lightning-asr-xxx"} 3

# HELP asr_total_requests Total requests processed
# TYPE asr_total_requests counter
asr_total_requests{pod="lightning-asr-xxx"} 1523

...

Test Prometheus Query

Access Prometheus UI and test queries:

asr_active_requests
rate(asr_total_requests[5m])
histogram_quantile(0.95, asr_request_duration_seconds_bucket)

Check Prometheus Targets

kubectl port-forward -n default svc/smallest-prometheus-stack-prometheus 9090:9090

Navigate to: http://localhost:9090/targets

Verify Lightning ASR targets are “UP”

View Prometheus Logs

kubectl logs -n default -l app.kubernetes.io/name=prometheus --tail=100

Look for scrape errors.

Troubleshooting

Metrics Not Appearing

Check ServiceMonitor is created:

kubectl get servicemonitor -n smallest

Check Prometheus is discovering:

kubectl logs -n default -l app.kubernetes.io/name=prometheus | grep lightning-asr

Check service has metrics port:

kubectl get svc lightning-asr -n smallest -o yaml

Should show:

ports:
  - name: metrics
    port: 2269

Custom Metrics Not Available

Check Prometheus Adapter logs:

kubectl logs -n kube-system -l app.kubernetes.io/name=prometheus-adapter

Verify adapter configuration:

kubectl get configmap prometheus-adapter -n kube-system -o yaml

Test API manually:

kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1" | jq .

High Cardinality Issues

If Prometheus is using too much memory:

1. Reduce label cardinality
2. Increase retention limits
3. Use recording rules for complex queries

kube-prometheus-stack:
  prometheus:
    prometheusSpec:
      resources:
        requests:
          memory: 4Gi
        limits:
          memory: 8Gi

Best Practices

Use Recording Rules

Pre-compute expensive queries:

- record: asr:requests:rate5m
  expr: rate(asr_total_requests[5m])

Then use in HPA instead of raw query

Set Appropriate Scrape Intervals

Balance responsiveness vs storage:

• Fast autoscaling: 15s
• Normal: 30s
• Cost-optimized: 60s

Enable Persistence

Always persist Prometheus data:

storageSpec:
  volumeClaimTemplate:
    spec:
      resources:
        requests:
          storage: 100Gi

Monitor Prometheus Itself

Track Prometheus performance:

• Query duration
• Scrape duration
• Memory usage
• TSDB size

Use Grafana for Visualization

Don’t rely on Prometheus UI

Use Grafana dashboards for ops

See Grafana Dashboards

What’s Next?

HPA Configuration

Use metrics for autoscaling

Grafana Dashboards

Visualize metrics