AI Intermediate: Infrastructure & Agentic SystemsDay 12 of 30 · Week 2: Kubernetes Foundation for MLOps
Course outline▾

Week 3 · High-Performance Networking & Security Mesh

  1. Day 15AI Network Bottlenecks: Transitioning from iptables to eBPF
  2. Day 16Cilium & eBPF: Sub-Millisecond Routing for Heavy AI Microservices
  3. Day 17Service Mesh 101: Installing Istio over Your Local Cluster
  4. Day 18Envoy Proxy as an AI Gateway: Load Balancing and Terminating LLM Prompts
  5. Day 19Zero-Trust mTLS: Encrypting Traffic Between Autonomous Agents
  6. Day 20API Rate Limiting & Egress Control: Preventing Agents from Exfiltrating Data
  7. Day 21Distributed Tracing: Debugging Agent Workflows with Jaeger and OpenTelemetry

Week 4 · Enterprise Agents & CI/CD Workflows

  1. Day 22Introduction to MLOps: Bridging AI Development with Continuous Integration
  2. Day 23GitOps with ArgoCD: Automated, Self-Healing AI Deployments
  3. Day 24Building Ephemeral AI Sandboxes: Secure Execution Environments for Agent Code
  4. Day 25Multi-Agent Orchestration Frameworks: Exploring Mastra, LangGraph, and CrewAI
  5. Day 26Tool-Calling in Production: Connecting Agents Securely to Internal Databases
  6. Day 27MLOps Guardrails: Evaluating Agents and Preventing Hallucinations in Real-Time
  7. Day 28Continuous Training Pipelines: Connecting Data Ingestion to Automated Model Retraining
  8. Day 29Managing Technical Debt and Infrastructure Drift in Enterprise AI
  9. Day 30The AI-Native Platform: Scaling from a Sandbox Environment to Global Cloud Architecture
Open the course page →

Day 12: Auto-Scaling AI Inference — Using Horizontal Pod Autoscalers for Spiky API Traffic

2026-11-18 · 8 min read

AI traffic is volatile, and idle GPU or high-memory instances are expensive. You want replicas to follow demand. The obvious tool is the Horizontal Pod Autoscaler (HPA), but its default signals, CPU and memory, mislead you for inference servers.

The pipeline

SCALE ON WHAT USERS FEEL: THE REQUEST QUEUERequestsGatewayvLLMExporterPrometheusAdapterHPAqueue above ~10 waiting → HPA scales out, for example by +2 podsExposing a custom metric to the HPA needs Prometheus Adapter (or KEDA).
FIG 1 · The inference engine exports queue depth, Prometheus stores it, an adapter exposes it to Kubernetes and the HPA adds or removes pods based on it.

Requests reach the inference engine, which exports a queue metric. Prometheus scrapes it, an adapter exposes it to Kubernetes as a custom metric and the HPA adds or removes pods based on it.

Why CPU and memory fail

WHY CPU AND MEMORY MISLEAD AN AI AUTOSCALERMEMORYthe engine reserves 85 to 90%of GPU memory at boothigh even at zero traffica memory HPA would max outCPU / GPU USEone request can drive thedevice near 100%continuous batching maystill have spare capacityso the signal is noisyBetter signals: waiting-queue depth, or time per output token (TPOT).They track the user's wait, which is what scaling should protect.
FIG 2 · Inference engines pre-allocate memory and saturate devices even under light load, so CPU and memory utilisation say little about demand. Queue depth and time per output token track what users experience.
  • Memory. Runtimes such as vLLM reserve 85 to 90% of GPU memory at start-up for the weights and the KV-cache pool. That figure is high even with no traffic, so a memory-based rule would scale out to the maximum and stay there.
  • CPU and GPU use. A single request can push a device close to 100%, while continuous batching may still have room for many more. The signal is noisy.

A better signal is the depth of the waiting queue, or the time per output token (TPOT). Both track the wait users feel, which is what scaling should protect.

QUEUE DEPTH DRIVES THE REPLICA COUNT (ILLUSTRATIVE)0102030target 10 waiting per pod1112344432pods:waiting requests
FIG 3 · As the waiting queue grows, the HPA adds replicas, and it removes them again slowly once the queue drains. The numbers are illustrative.

The HPA manifest

This scales on the number of waiting requests per pod. vLLM exports this metric as vllm:num_requests_waiting, and Prometheus Adapter exposes it to Kubernetes under a name you configure, shown here as vllm_num_requests_waiting.

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: inference-engine-hpa
  namespace: ai-platform
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: inference-engine-gemma
  minReplicas: 1
  maxReplicas: 4
  metrics:
    - type: Pods
      pods:
        metric:
          name: vllm_num_requests_waiting
        target:
          type: AverageValue
          averageValue: "10"
  behavior:
    scaleUp:
      stabilizationWindowSeconds: 0
      policies:
        - { type: Percent, value: 100, periodSeconds: 15 }
    scaleDown:
      stabilizationWindowSeconds: 300
      policies:
        - { type: Percent, value: 25, periodSeconds: 60 }

Exposing a custom metric to the HPA needs an adapter such as Prometheus Adapter, or you can use KEDA, which can scale directly on a Prometheus query. Check the exact metric name your vLLM version exports.

Scale up fast, scale down slowly

SCALE UP FAST, SCALE DOWN SLOWLYscaleUpstabilizationWindow: 0 sup to +100% every 15 sreact to a spike at once(new pods still need load time)scaleDownstabilizationWindow: 300 sat most -25% every 60 sdo not throw away warm podsthat took minutes to loadminReplicas 1, maxReplicas 4, target averageValue 10 on the queue metric.A GPU pod is slow and costly to start, so flapping is expensive.
FIG 4 · Scale out aggressively to absorb a spike, and scale in cautiously, because a model pod takes minutes to load and flapping wastes that work.

A new model pod takes minutes to load, so you want to react to a spike immediately and avoid throwing away warm pods the moment traffic dips. The behavior block encodes that: no stabilisation delay when scaling up, and a five-minute window with gentle steps when scaling down.

Scaling on the queue ties your replica count to the user experience, and it keeps expensive capacity from sitting idle.

Run it yourself

This is the exact example from the video. A small Python model of the autoscaler's rule, using the lesson's numbers. It is not the real HPA.

import math

target, low, high = 10, 1, 4   # waiting requests per pod
window = 5                     # minutes, like 300 seconds
recs = []

for minute, waiting in enumerate([5, 30, 60, 20, 5, 5, 5, 5]):
    want = math.ceil(waiting / target)
    recs.append(min(high, max(low, want)))
    pods = max(recs[-window:])
    print(f"min {minute}: waiting {waiting:2d} -> {pods} pods")

Run it:

python hpa.py

Output I got while recording the video:

min 0: waiting  5 -> 1 pods
min 1: waiting 30 -> 3 pods
min 2: waiting 60 -> 4 pods
min 3: waiting 20 -> 4 pods
min 4: waiting  5 -> 4 pods
min 5: waiting  5 -> 4 pods
min 6: waiting  5 -> 4 pods
min 7: waiting  5 -> 2 pods

Coming Up Next

Day 13: Helm charts: packaging, versioning and updating AI deployments.

#HPA #Autoscaling #MLOps #Prometheus #CloudCostOptimization #SRE #Kubernetes