{"id":54055,"date":"2025-11-12T08:59:58","date_gmt":"2025-11-12T08:59:58","guid":{"rendered":"https:\/\/www.devopsschool.com\/blog\/?p=54055"},"modified":"2026-02-21T08:28:46","modified_gmt":"2026-02-21T08:28:46","slug":"kubernetes-tutorials-traffic-driven-autoscaling-on-kubernetes-comparing-keda-hpa-vpa-custom-adapters","status":"publish","type":"post","link":"https:\/\/www.devopsschool.com\/blog\/kubernetes-tutorials-traffic-driven-autoscaling-on-kubernetes-comparing-keda-hpa-vpa-custom-adapters\/","title":{"rendered":"Kubernetes Tutorials: Traffic-Driven Autoscaling on Kubernetes & Comparing KEDA, HPA, VPA & Custom Adapters"},"content":{"rendered":"\n

Comparing KEDA, HPA, VPA & Custom Adapters for Real-World Scaling with Cost, Complexity & Best Practices<\/em><\/h3>\n\n\n\n
Category<\/strong><\/th>KEDA (Event-driven)<\/strong><\/th>Prometheus-based (Adapter\/HPA)<\/strong><\/th>Datadog-based (Cluster Agent)<\/strong><\/th>CloudWatch-based (Adapter\/HPA)<\/strong><\/th><\/tr><\/thead>
Primary Function<\/strong><\/td>Event-driven autoscaler (creates HPAs dynamically from external metrics like ALB, SQS, Kafka, etc.)<\/td>Uses HPA with Prometheus metrics via adapter (e.g., kube-metrics-adapter or prometheus-adapter)<\/td>Uses Datadog metrics (via Cluster Agent) as external metrics for HPA<\/td>Uses CloudWatch metrics (via AWS CloudWatch Metrics Adapter) for HPA<\/td><\/tr>
Metric Source<\/strong><\/td>Multiple external sources: CloudWatch, Prometheus, SQS, Kafka, HTTP, etc. (50+ scalers)<\/td>Prometheus time-series metrics (scraped from exporters or apps)<\/td>Datadog platform metrics (ingested from AWS, custom apps, APM)<\/td>AWS CloudWatch (e.g., ALB metrics, RDS, SQS, Lambda, etc.)<\/td><\/tr>
Data Flow Model<\/strong><\/td>Pull metrics or events \u2192 internal HPA \u2192 scale<\/td>Prometheus scrapes \u2192 adapter \u2192 Kubernetes Metrics API \u2192 HPA<\/td>Datadog agent \u2192 Cluster Agent \u2192 External Metrics API \u2192 HPA<\/td>CloudWatch Adapter \u2192 External Metrics API \u2192 HPA<\/td><\/tr>
Setup Complexity<\/strong><\/td>\ud83d\udfe2 Medium (Helm + few YAMLs; no exporter needed)<\/td>\ud83d\udd35 Medium-High (need Prometheus + adapter configuration)<\/td>\ud83d\udfe3 Medium (if Datadog is already deployed)<\/td>\ud83d\udfe0 Medium (adapter installation + IAM + mappings)<\/td><\/tr>
Integration with ALB Traffic<\/strong><\/td>\u2705 Native (via CloudWatch scaler \u2013 uses RequestCountPerTarget<\/code>, TargetResponseTime<\/code>)<\/td>\u26a0\ufe0f Requires Prometheus CloudWatch exporter (YACE or similar)<\/td>\u2705 Native (Datadog already pulls ALB metrics)<\/td>\u2705 Native (direct access to ALB metrics)<\/td><\/tr>
Supports Scale-to-Zero<\/strong><\/td>\u2705 Yes<\/td>\u274c No (HPA cannot scale to zero)<\/td>\u274c No<\/td>\u274c No<\/td><\/tr>
Responsiveness \/ Latency<\/strong><\/td>~30\u201360 seconds (depends on CloudWatch polling)<\/td>~15\u201330 seconds (depends on scrape interval)<\/td>~30\u201360 seconds (depends on Datadog ingestion)<\/td>~60 seconds (CloudWatch metric delay)<\/td><\/tr>
Operational Cost<\/strong><\/td>\ud83d\udcb2 Low (CloudWatch API calls only)<\/td>\ud83d\udcb2\ud83d\udcb2 Medium (Prometheus infra + storage + exporter costs)<\/td>\ud83d\udcb2\ud83d\udcb2\ud83d\udcb2 High (Datadog licensing per host\/container)<\/td>\ud83d\udcb2 Low (CloudWatch API calls)<\/td><\/tr>
Infrastructure Overhead<\/strong><\/td>Lightweight (1 KEDA controller)<\/td>Heavy (Prometheus, exporters, adapter)<\/td>Moderate (Datadog Cluster Agent)<\/td>Moderate (Adapter deployment)<\/td><\/tr>
Ease of Maintenance<\/strong><\/td>\ud83d\udfe2 Easy \u2013 one Helm upgrade for all namespaces<\/td>\ud83d\udd35 Moderate \u2013 maintain adapter & Prometheus<\/td>\ud83d\udfe3 Easy if Datadog already managed<\/td>\ud83d\udfe0 Moderate \u2013 periodic IAM & adapter updates<\/td><\/tr>
EKS Auto Mode Compatibility<\/strong><\/td>\u2705 Fully compatible \u2013 scales pods; NodePools handle nodes<\/td>\u2705 Compatible<\/td>\u2705 Compatible<\/td>\u2705 Compatible<\/td><\/tr>
Multi-Namespace Scaling<\/strong><\/td>\u2705 Native support (Scoped per namespace)<\/td>\u2705 Supported<\/td>\u2705 Supported<\/td>\u2705 Supported<\/td><\/tr>
Security \/ IAM<\/strong><\/td>Uses IRSA or static keys for AWS APIs<\/td>No AWS permissions required (depends on Prometheus)<\/td>Uses Datadog API key & IAM integration<\/td>Uses IRSA for AWS CloudWatch read access<\/td><\/tr>
Supported Triggers \/ Metrics<\/strong><\/td>50+ sources (CloudWatch, Kafka, RabbitMQ, HTTP, Redis, MySQL, etc.)<\/td>Limited to Prometheus metrics<\/td>Limited to Datadog metrics<\/td>Limited to AWS metrics<\/td><\/tr>
Scales on Events (not metrics)<\/strong><\/td>\u2705 Yes<\/td>\u274c No<\/td>\u274c No<\/td>\u274c No<\/td><\/tr>
Can Combine Multiple Triggers<\/strong><\/td>\u2705 Yes (multi-trigger scaling rules)<\/td>\u26a0\ufe0f Only via complex PromQL expressions<\/td>\u26a0\ufe0f Limited (Datadog composite metrics)<\/td>\u26a0\ufe0f Limited (one metric per HPA)<\/td><\/tr>
Recommended For<\/strong><\/td>Event-driven \/ traffic-based workloads (ALB, queues, web APIs)<\/td>Resource or app-metric-based workloads<\/td>Organizations using Datadog for monitoring & APM<\/td>AWS-centric workloads without Prometheus<\/td><\/tr>
Learning Curve<\/strong><\/td>\ud83d\udfe2 Low<\/td>\ud83d\udd35 Medium<\/td>\ud83d\udfe3 Low<\/td>\ud83d\udfe0 Medium<\/td><\/tr>
Vendor Lock-in<\/strong><\/td>Low (Open Source)<\/td>Low (OSS ecosystem)<\/td>High (Datadog SaaS)<\/td>Medium (AWS-only)<\/td><\/tr>
Community & Ecosystem<\/strong><\/td>Very active (CNCF Sandbox Project)<\/td>Large (K8s ecosystem standard)<\/td>Proprietary (Datadog documentation)<\/td>AWS-maintained (moderate community)<\/td><\/tr>
Use with WAF + ALB<\/strong><\/td>\u2705 Seamless (uses ALB TG metrics directly)<\/td>\u26a0\ufe0f Need exporter for ALB metrics<\/td>\u2705 Seamless (Datadog ALB integration)<\/td>\u2705 Seamless (ALB metrics native in CloudWatch)<\/td><\/tr>
Example Metric<\/strong><\/td>CloudWatch \u2192 RequestCountPerTarget<\/code>, TargetResponseTime<\/code><\/td>Prometheus \u2192 nginx_ingress_controller_requests_total<\/code><\/td>Datadog \u2192 aws.applicationelb.request_count<\/code><\/td>CloudWatch \u2192 RequestCountPerTarget<\/code><\/td><\/tr>
Scale Behavior Visualization<\/strong><\/td>KEDA Metrics API + Grafana dashboards<\/td>Prometheus \/ Grafana<\/td>Datadog Dashboards<\/td>CloudWatch Dashboards<\/td><\/tr>
Maturity (as of 2026)<\/strong><\/td>\u2b50\u2b50\u2b50\u2b50\u2b50 (CNCF Incubating)<\/td>\u2b50\u2b50\u2b50\u2b50<\/td>\u2b50\u2b50\u2b50\u2b50<\/td>\u2b50\u2b50\u2b50\u2b50<\/td><\/tr>
Overall Recommendation (for EKS + ALB)<\/strong><\/td>\u2705\u2705\u2705 Best Option<\/strong><\/td>\u2705 Good (if Prometheus is already in place)<\/td>\u2699\ufe0f Suitable for Datadog-native orgs<\/td>\u2705 Good fallback if KEDA not allowed<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\n\n\n\n

\ud83c\udf10 1 | Architecture Overview<\/h2>\n\n\n\n

Flow:<\/strong>
Client \u2192 DNS \u2192 WAF \u2192 ALB \u2192 TargetGroup \u2192 EKS Service\/Pods<\/code><\/p>\n\n\n\n

Layer<\/th>Purpose<\/th>Key AWS \/ K8s Component<\/th><\/tr><\/thead>
Edge Security<\/strong><\/td>Filter malicious traffic<\/td>AWS WAF (Web ACL)<\/td><\/tr>
Load Balancing<\/strong><\/td>Distribute inbound requests<\/td>ALB (AWS Load Balancer Controller)<\/td><\/tr>
Routing<\/strong><\/td>Path\/host-based dispatch to namespaces<\/td>Kubernetes Ingress<\/td><\/tr>
Compute<\/strong><\/td>Run workloads<\/td>EKS Pods\/Deployments<\/td><\/tr>
Node Capacity<\/strong><\/td>Provision nodes automatically<\/td>EKS Auto Mode NodePools (Karpenter)<\/td><\/tr>
Autoscaling Brain<\/strong><\/td>Adjust replicas dynamically<\/td>KEDA \/ HPA \/ VPA \/ Custom Adapter<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

With EKS Auto Mode<\/strong>, AWS manages node scaling.
Your responsibility is pod-level scaling<\/strong> \u2014 deciding how many replicas each service needs<\/em> based on traffic or resource metrics.<\/p>\n\n\n\n

\n\n\n\n

\ud83e\udde9 2 | Namespace-Scoped Design Pattern<\/h2>\n\n\n\n
    \n
  • Each microservice (e.g., booking<\/code>, auth<\/code>, medical<\/code>, telematics<\/code>) lives in its own namespace<\/strong>.<\/li>\n\n\n\n
  • Each namespace has its Ingress, Service, Deployment, ConfigMap, and autoscaler<\/strong> objects.<\/li>\n\n\n\n
  • Optionally, multiple namespaces can share one ALB<\/strong> via alb.ingress.kubernetes.io\/group.name<\/code> to save cost while keeping per-namespace isolation.<\/li>\n<\/ul>\n\n\n\n
    \n\n\n\n

    \u2699\ufe0f 3 | Ingress & WAF Setup (Shared ALB Example)<\/h2>\n\n\n
    apiVersion: networking.k8s.io\/v1\nkind: Ingress\nmetadata:\n  name: booking-ing\n  namespace: booking\n  annotations:\n    kubernetes.io\/ingress.class: alb\n    alb.ingress.kubernetes.io\/group.name: shared-edge\n    alb.ingress.kubernetes.io\/group.order: \"20\"\n    alb.ingress.kubernetes.io\/target-type: ip\n    alb.ingress.kubernetes.io\/scheme: internet-facing\n    alb.ingress.kubernetes.io\/wafv2-acl-arn: arn:aws:wafv2:ap-northeast-1:111111111111:regional\/webacl\/mywebacl\/abcd1234\nspec:\n  rules:\n  - host: api.example.com\n    http:\n      paths:\n      - path: \/booking\n        pathType: Prefix\n        backend:\n          service:\n            name: booking-svc\n            port:\n              number: 80\n<\/code><\/span><\/pre>\n\n\n
    Each namespace can repeat this pattern using different paths (\/auth<\/code>, \/legal<\/code>, etc.) but share the same group.name<\/code> \u2192 one ALB under one WAF.<\/p>\n\n\n\n
    \n\n\n\n
    \ud83d\ude80 4 | Autoscaling Options for Pod Level Control<\/h2>\n\n\n\n
    Below are five viable mechanisms<\/strong> for pod autoscaling inside EKS.<\/p>\n\n\n\n
    #<\/th>Method<\/th>Scaling Source<\/th>Scales To Zero<\/th>Works with ALB Metrics<\/th>Typical Latency<\/th>Setup Time<\/th>Maint. Effort<\/th>Approx. Cost*<\/th>Skill Level<\/th><\/tr><\/thead>
    1<\/td>KEDA<\/strong><\/td>External events (CloudWatch ALB, SQS, Prometheus, etc.)<\/td>\u2705<\/td>\u2705 (native scaler)<\/td>30-60 s<\/td>\u2699\ufe0f Medium<\/td>\ud83e\udde9 Low (once installed)<\/td>\ud83d\udcb2\ud83d\udcb2 CloudWatch API calls<\/td>Intermediate<\/td><\/tr>
    2<\/td>HPA<\/strong><\/td>CPU \/ memory \/ custom metrics<\/td>\u274c<\/td>\u26a0\ufe0f via adapter<\/td>15-30 s<\/td>\u2699\ufe0f Low<\/td>\ud83e\udde9 Low<\/td>\ud83d\udcb2 free<\/td>Beginner<\/td><\/tr>
    3<\/td>VPA<\/strong><\/td>Internal resource usage<\/td>\u274c<\/td>\u274c<\/td>N\/A<\/td>\u2699\ufe0f Medium<\/td>\ud83e\udde9 Low<\/td>\ud83d\udcb2 free<\/td>Intermediate<\/td><\/tr>
    4<\/td>Custom Metric Adapter<\/strong><\/td>Prometheus \/ CloudWatch<\/td>\u274c<\/td>\u2705 with manual mapping<\/td>45-60 s<\/td>\u2699\ufe0f High<\/td>\ud83e\udde9 High<\/td>\ud83d\udcb2\ud83d\udcb2 metrics infra<\/td>Advanced<\/td><\/tr>
    5<\/td>Manual Scaling<\/strong><\/td>Human input<\/td>\u274c<\/td>\u274c<\/td>N\/A<\/td>\u2699\ufe0f Instant<\/td>\ud83e\udde9 High Opex<\/td>\ud83d\udcb2 none<\/td>Basic<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    * Cost = relative AWS service charges + operational overhead<\/p>\n\n\n\n
    \n\n\n\n
    \ud83e\uddee 5 | Detailed Analysis of Each Approach<\/h2>\n\n\n\n
    \ud83d\udd39 A | KEDA (Event-Driven Autoscaler)<\/h3>\n\n\n\n
    How it works:<\/strong>
    KEDA reads external metrics (CloudWatch ALB RequestCountPerTarget, TargetResponseTime, SQS depth, PromQL queries, etc.) and creates an internal HPA.<\/p>\n\n\n\n
    Pros<\/strong><\/p>\n\n\n\n