GPU Autoscaling for AI Workloads: HPA, KEDA, Pre-Warming

GPU instances for AI are expensive: an A100 hour costs a lot, and idle time eats your budget. Cold start of a new pod takes 3–10 minutes (model loading, CUDA initialization), and standard CPU autoscaling does not account for GPU specifics. Without proper autoscaling, you pay for idle time or lose requests during peaks. We solve this: we configure scaling that actually meets SLA and reduces costs by 30–40%.

Problems We Solve

Cold start: GPU pod startup time 3–10 minutes. During this, the request queue overflows. Solutions: keepalive pods (minimum 1), pre-warming (start at 70% queue load), and buffering via request queue. Using spot instances worsens the problem — they can be interrupted anytime, so we combine spot with on-demand for critical services.

GPU utilization vs queue depth: GPU load during a long request is 100%, but new requests wait. Relying on utilization is a mistake. The correct metric is vllm_num_requests_waiting or queue_depth. We use a combination: scale-up by queue depth, scale-down by utilization.

Thrashing pods: Frequent spin-ups and downs due to sharp spikes. Configuring stabilization windows prevents this: scale-down after 10 minutes, scale-up within 30 seconds.

Why Queue Depth Is the Main Metric for LLM?

GPU utilization does not reflect queue latency. When a model processes a long request, GPU is 100% busy, but new requests are queued. Queue depth shows the real resource need. It should be the primary trigger for scale-up.

How We Configure GPU Autoscaling

Kubernetes HPA with Custom Metrics

# Prometheus Adapter for custom metrics
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: vllm-autoscaler
  namespace: ai-serving
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: vllm-llama3
  minReplicas: 1
  maxReplicas: 8
  metrics:
    # Main metric: queue of waiting requests
    - type: Pods
      pods:
        metric:
          name: vllm_pending_requests
        target:
          type: AverageValue
          averageValue: "5"          # scale when > 5 requests in queue per pod

    # Additional: GPU utilization (for scale-down)
    - type: Pods
      pods:
        metric:
          name: nvidia_gpu_duty_cycle
        target:
          type: AverageValue
          averageValue: "70"         # scale-down when < 70% utilization

  behavior:
    scaleUp:
      stabilizationWindowSeconds: 30     # fast scale-up
      policies:
        - type: Pods
          value: 2
          periodSeconds: 60              # +2 pods every minute
    scaleDown:
      stabilizationWindowSeconds: 600    # slow scale-down (10 minutes)
      policies:
        - type: Pods
          value: 1
          periodSeconds: 300             # -1 pod every 5 minutes

KEDA for Event-Driven Autoscaling

KEDA is more flexible than HPA: scaling by Prometheus, Kafka, RabbitMQ, SQS. Below is an example configuration.

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: vllm-keda-scaler
  namespace: ai-serving
spec:
  scaleTargetRef:
    name: vllm-llama3
  minReplicaCount: 1
  maxReplicaCount: 10
  cooldownPeriod: 300
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring.svc.cluster.local:9090
        metricName: vllm_queue_size
        query: sum(vllm_num_requests_waiting{namespace="ai-serving"})
        threshold: "10"           # 1 replica per 10 waiting requests

    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring.svc.cluster.local:9090
        metricName: request_rate
        query: rate(http_requests_total{job="vllm"}[2m])
        threshold: "20"           # additional trigger by RPS

Cloud-Native Autoscaling

For cloud GPU instances, we use Auto Scaling Groups with custom metrics:

import boto3

autoscaling = boto3.client('autoscaling', region_name='us-east-1')

autoscaling.put_scaling_policy(
    AutoScalingGroupName='llm-gpu-asg',
    PolicyName='scale-on-queue-depth',
    PolicyType='TargetTrackingScaling',
    TargetTrackingConfiguration={
        'CustomizedMetricSpecification': {
            'MetricName': 'LLMQueueDepth',
            'Namespace': 'Custom/LLMMetrics',
            'Statistic': 'Average',
        },
        'TargetValue': 5.0,
        'ScaleInCooldown': 300,
        'ScaleOutCooldown': 60,
        'DisableScaleIn': False,
    }
)

Publishing Custom Metrics from vLLM

We collect metrics directly from the inference server:

from prometheus_client import Gauge, start_http_server
import requests
import time

QUEUE_SIZE = Gauge('llm_queue_depth', 'Number of pending requests')
GPU_MEMORY = Gauge('llm_gpu_memory_used_gb', 'GPU memory usage in GB', ['gpu_id'])

def collect_metrics():
    response = requests.get("http://localhost:8000/metrics").text
    for line in response.split('\n'):
        if 'vllm:num_requests_waiting' in line and not line.startswith('#'):
            queue_size = float(line.split()[-1])
            QUEUE_SIZE.set(queue_size)

    import subprocess
    result = subprocess.run(
        ['nvidia-smi', '--query-gpu=memory.used', '--format=csv,noheader,nounits'],
        capture_output=True, text=True
    )
    for i, mem_mb in enumerate(result.stdout.strip().split('\n')):
        GPU_MEMORY.labels(gpu_id=str(i)).set(float(mem_mb) / 1024)

start_http_server(9091)
while True:
    collect_metrics()
    time.sleep(15)

How Pre-Warming Prevents Cold Start?

We prevent cold start by forecasting load. If queue depth exceeds 70% of the maximum, we launch an additional pod in advance. Strategy code:

class PreWarmingStrategy:
    def __init__(self, warmup_threshold: float = 0.7, warmup_lead_time: int = 180):
        self.warmup_threshold = warmup_threshold
        self.warmup_lead_time = warmup_lead_time

    def should_scale_up(self, current_queue: int, max_queue: int, forecast: list) -> bool:
        if current_queue / max_queue >= self.warmup_threshold:
            return True
        future_queue = forecast[self.warmup_lead_time // 15]
        return future_queue / max_queue >= self.warmup_threshold

Why GPU Autoscaling Is Harder Than CPU?

Key differences: GPU instances cannot be "split" between services, cold start is much longer (model loading ~5 GB), and load metrics are misleading. On CPU you rely on CPU utilization, on GPU — only on queue depth and request arrival rate. Also, the cost of error is higher: an extra GPU pod means extra monthly expenses.

How to Choose a Metric for Scaling?

The best solution is a combination: queue depth for scale-up and GPU utilization for scale-down.

What Our Work Includes

We offer turnkey implementation:

1. Infrastructure audit — load analysis, bottleneck identification, instance selection.
2. Scaling scheme design — metric selection, HPA/KEDA configuration, behavior optimization (stabilization windows, policies).
3. Implementation — Prometheus stack deployment, vLLM/TGI integration, cloud autoscaling group setup.
4. Testing — load testing with peak simulation, threshold calibration.
5. Documentation and training — runbook for on-call, metric and alert description.
6. Post-release support — monitoring for first weeks, policy adjustments based on actuals.

Timeline and Savings

Estimated timeline:

• Basic setup (metrics + HPA) — from 5 days.
• Extended configuration (KEDA + pre-warming) — from 10 days.
• Full cycle with load testing — from 3 weeks.

Cost is calculated individually based on stack complexity and workload. Our clients save 30–40% monthly GPU costs after implementation.

Order GPU autoscaling implementation and start saving within a week. We have been working with AI infrastructure for 5+ years, completed 20+ projects. We guarantee stable operation — we include SLA on scaling response time in the contract. Kubernetes HPA and KEDA are proven tools. Contact us for a consultation on your project — we will analyze your load and offer the optimal scheme.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.