Module 5.4: Model Serving & Inference

Discipline Track | Complexity: [COMPLEX] | Time: 40-45 min

Prerequisites

Before starting this module:

• Module 5.3: Model Training & Experimentation
• Understanding of REST APIs
• Basic Kubernetes concepts (Deployments, Services)
• Familiarity with containerization (Docker)

What You’ll Be Able to Do

After completing this module, you will be able to:

• Implement model serving infrastructure using KServe, Seldon Core, or Triton on Kubernetes
• Design inference architectures that handle model versioning, A/B testing, and canary rollouts
• Configure autoscaling policies for model endpoints based on request latency and throughput SLOs
• Build model serving pipelines with preprocessing, inference, and postprocessing stages

Why This Module Matters

A model in a notebook isn’t a product. The best fraud detection model is worthless if it takes 30 seconds to return a prediction. Users won’t wait, transactions will fail, business will suffer.

Model serving is where ML meets reality. Latency matters. Scalability matters. Reliability matters. A prediction service that’s down loses money—or worse, defaults to allowing fraudulent transactions.

The gap between “model works in Jupyter” and “model serves 10,000 requests per second” is where most ML projects fail.

Did You Know?

• Netflix serves 150 million predictions per second using a combination of batch and real-time serving—different use cases need different architectures
• The latency budget for real-time inference is often <100ms—including network, feature lookup, and model computation. Every millisecond counts
• Google found that a 0.5-second delay in search results reduces traffic by 20%. For ML predictions embedded in user flows, latency is everything
• Model serving costs can exceed training costs by 10x—training happens once, serving happens continuously. Optimize for inference, not just training

Serving Patterns

Different use cases require different serving patterns:

┌─────────────────────────────────────────────────────────────────┐
│                    MODEL SERVING PATTERNS                        │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ONLINE (Real-time)                                              │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │ Request ──▶ Feature Lookup ──▶ Model ──▶ Response        │   │
│  │                                                          │   │
│  │ Latency: <100ms  │  Throughput: High  │  Example: Fraud  │   │
│  └──────────────────────────────────────────────────────────┘   │
│                                                                  │
│  BATCH (Scheduled)                                               │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │ Data ──▶ Model ──▶ Predictions ──▶ Store                 │   │
│  │                                                          │   │
│  │ Latency: Hours   │  Throughput: Very High  │  Example:   │   │
│  │                                               Recommendations│
│  └──────────────────────────────────────────────────────────┘   │
│                                                                  │
│  STREAMING (Event-driven)                                        │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │ Event Stream ──▶ Model ──▶ Event Stream                  │   │
│  │                                                          │   │
│  │ Latency: Seconds │  Throughput: High  │  Example: Anomaly│   │
│  └──────────────────────────────────────────────────────────┘   │
│                                                                  │
│  EDGE (On-device)                                                │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │ Device Input ──▶ Local Model ──▶ Prediction              │   │
│  │                                                          │   │
│  │ Latency: <10ms  │  Throughput: N/A    │  Example: Face ID│   │
│  └──────────────────────────────────────────────────────────┘   │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Choosing a Pattern

Use Case	Pattern	Why
Fraud detection	Online	Must block transaction in real-time
Product recommendations	Batch	Pre-compute, serve from cache
Log anomaly detection	Streaming	Process events as they arrive
Mobile image classification	Edge	Works offline, low latency
Credit scoring	Online + Batch	Real-time for decisions, batch for reports

War Story: The Recommendation Catastrophe

A startup built an amazing recommendation model. In tests, it was perfect—personalized, relevant, engaging. They deployed it as a real-time service.

Launch day: 500ms latency per recommendation. Pages with 20 recommendations took 10 seconds to load. Users bounced. Revenue tanked.

The fix? Batch predictions. Pre-compute recommendations hourly, serve from Redis. Latency dropped to 5ms. Same model, different serving pattern.

Online Serving Architecture

┌─────────────────────────────────────────────────────────────────┐
│                  ONLINE SERVING ARCHITECTURE                     │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  Client Request                                                  │
│       │                                                          │
│       ▼                                                          │
│  ┌─────────────┐                                                │
│  │   Gateway   │  (Rate limiting, auth, routing)                │
│  │  (Kong/Istio)│                                                │
│  └──────┬──────┘                                                │
│         │                                                        │
│         ▼                                                        │
│  ┌─────────────┐     ┌─────────────┐                            │
│  │  Inference  │◀────│   Feature   │  (Low-latency lookup)      │
│  │   Service   │     │    Store    │                            │
│  │  (KServe)   │     │   (Redis)   │                            │
│  └──────┬──────┘     └─────────────┘                            │
│         │                                                        │
│         ▼                                                        │
│  ┌─────────────┐     ┌─────────────┐                            │
│  │   Model     │     │  Monitoring │  (Latency, throughput)     │
│  │  (loaded)   │     │ (Prometheus)│                            │
│  └──────┬──────┘     └─────────────┘                            │
│         │                                                        │
│         ▼                                                        │
│  Response                                                        │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Model Serving Frameworks

Framework Comparison

Framework	Best For	Pros	Cons
KServe	Kubernetes-native ML	Autoscaling, multi-framework	K8s complexity
Seldon Core	Enterprise ML	Explainability, A/B testing	Complex setup
BentoML	Simple deployment	Easy packaging	Less scalable
TorchServe	PyTorch models	Optimized for PyTorch	Framework-specific
TensorFlow Serving	TensorFlow models	High performance	TF only
Triton	Multi-framework GPU	Best GPU utilization	NVIDIA only

KServe (Kubernetes Serving)

KServe (formerly KFServing) is the standard for Kubernetes-native model serving:

┌─────────────────────────────────────────────────────────────────┐
│                         KSERVE                                   │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ┌─────────────────────────────────────────────────────────┐    │
│  │                 InferenceService                         │    │
│  │  ┌───────────────┐  ┌───────────────┐  ┌─────────────┐  │    │
│  │  │  Transformer  │  │   Predictor   │  │  Explainer  │  │    │
│  │  │  (preprocess) │  │   (model)     │  │ (optional)  │  │    │
│  │  └───────┬───────┘  └───────┬───────┘  └──────┬──────┘  │    │
│  │          │                  │                 │          │    │
│  │          └──────────────────┼─────────────────┘          │    │
│  │                             │                            │    │
│  │  ┌──────────────────────────▼────────────────────────┐  │    │
│  │  │                    Knative                         │  │    │
│  │  │  • Scale to zero    • Canary deployments          │  │    │
│  │  │  • Auto-scaling     • Traffic splitting           │  │    │
│  │  └───────────────────────────────────────────────────┘  │    │
│  └─────────────────────────────────────────────────────────┘    │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

KServe InferenceService

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: fraud-detector
  namespace: ml-serving
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://models/fraud-detector/v1
      resources:
        requests:
          cpu: "500m"
          memory: "512Mi"
        limits:
          cpu: "1"
          memory: "1Gi"
    minReplicas: 1
    maxReplicas: 10
    scaleTarget: 10  # Concurrent requests per replica
    scaleMetric: concurrency

Deploying to KServe

# Apply InferenceService
kubectl apply -f inference-service.yaml

# Check status
kubectl get inferenceservice fraud-detector -n ml-serving

# Get endpoint URL
kubectl get inferenceservice fraud-detector -n ml-serving -o jsonpath='{.status.url}'

# Test prediction
curl -X POST "http://${SERVICE_URL}/v1/models/fraud-detector:predict" \
  -H "Content-Type: application/json" \
  -d '{"instances": [[0.5, 0.3, 0.2, 0.8, 0.1]]}'

Deployment Strategies

Canary Deployments

CANARY DEPLOYMENT
─────────────────────────────────────────────────────────────────

Traffic Distribution:
┌───────────────────────────────────────────────────────────────┐
│                                                               │
│  Day 1:    v1 ████████████████████████████████████████ 100%  │
│            v2                                            0%   │
│                                                               │
│  Day 2:    v1 ████████████████████████████████████     90%   │
│            v2 ████                                       10%  │
│                                                               │
│  Day 3:    v1 ████████████████████                      50%   │
│            v2 ████████████████████                      50%   │
│                                                               │
│  Day 4:    v1                                            0%   │
│            v2 ████████████████████████████████████████ 100%  │
│                                                               │
└───────────────────────────────────────────────────────────────┘

Monitor: latency, errors, business metrics
Rollback: If metrics degrade, shift traffic back to v1

KServe Canary

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: fraud-detector
spec:
  predictor:
    # Primary (90% traffic)
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://models/fraud-detector/v1
    # Canary (10% traffic)
    canaryTrafficPercent: 10
  canary:
    predictor:
      model:
        modelFormat:
          name: sklearn
        storageUri: s3://models/fraud-detector/v2

Shadow Deployments

Shadow mode runs the new model on real traffic without affecting responses:

SHADOW DEPLOYMENT
─────────────────────────────────────────────────────────────────

User Request ──┬──▶ Production Model ──▶ User Response
               │
               └──▶ Shadow Model ──▶ Log Only (compare later)

Benefits:
• Real production traffic
• No user impact
• Compare performance on real data
• Validate before promotion

A/B Testing

For comparing model variants with statistical significance:

# Seldon Core A/B test
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: ab-test
spec:
  predictors:
    - name: model-a
      traffic: 50
      graph:
        name: classifier-a
        modelUri: gs://models/model-a
    - name: model-b
      traffic: 50
      graph:
        name: classifier-b
        modelUri: gs://models/model-b

Model Optimization

Why Optimize?

Metric	Impact of Optimization
Latency	50ms → 10ms (5x faster)
Throughput	100 RPS → 1000 RPS (10x higher)
Memory	1GB → 100MB (10x smaller)
Cost	$1000/mo → $100/mo (10x cheaper)

Optimization Techniques

MODEL OPTIMIZATION TECHNIQUES
─────────────────────────────────────────────────────────────────

QUANTIZATION (Reduce precision)
├── FP32 → FP16: 2x smaller, ~same accuracy
├── FP32 → INT8: 4x smaller, slight accuracy loss
└── Example: TensorRT, ONNX Runtime

PRUNING (Remove weights)
├── Unstructured: Zero out small weights
├── Structured: Remove entire neurons/layers
└── Trade-off: Size vs. accuracy

DISTILLATION (Smaller model learns from larger)
├── Teacher: Large, accurate model
├── Student: Small, fast model
└── Result: Student ~= Teacher accuracy, much faster

COMPILATION (Optimize for hardware)
├── ONNX: Standard format, cross-platform
├── TensorRT: NVIDIA GPU optimization
├── XLA: TensorFlow compilation
└── TorchScript: PyTorch JIT compilation

ONNX for Portability

# Convert sklearn model to ONNX
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType

# Define input type
initial_type = [('float_input', FloatTensorType([None, 4]))]

# Convert
onnx_model = convert_sklearn(sklearn_model, initial_types=initial_type)

# Save
with open("model.onnx", "wb") as f:
    f.write(onnx_model.SerializeToString())

# Run with ONNX Runtime
import onnxruntime as ort

session = ort.InferenceSession("model.onnx")
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name

predictions = session.run([output_name], {input_name: input_data})

Quantization Example

# TensorFlow Lite quantization
import tensorflow as tf

# Load model
model = tf.keras.models.load_model('model.h5')

# Convert with quantization
converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float16]  # or tf.int8

tflite_model = converter.convert()

# Save
with open('model_quantized.tflite', 'wb') as f:
    f.write(tflite_model)

# Size comparison
import os
print(f"Original: {os.path.getsize('model.h5') / 1024 / 1024:.2f} MB")
print(f"Quantized: {os.path.getsize('model_quantized.tflite') / 1024 / 1024:.2f} MB")

Scaling Strategies

Horizontal vs. Vertical Scaling

SCALING STRATEGIES
─────────────────────────────────────────────────────────────────

VERTICAL (Scale Up)
├── More CPU/memory per pod
├── Larger GPU
├── Limits: Max instance size
└── Use when: Model needs more resources

HORIZONTAL (Scale Out)
├── More replicas
├── Load balancing
├── Limits: Coordination overhead
└── Use when: Need more throughput

AUTO-SCALING
├── Scale on CPU/memory
├── Scale on requests per second
├── Scale on queue depth
└── Scale to zero (cost savings)

KServe Autoscaling

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: fraud-detector
  annotations:
    # Knative autoscaling
    autoscaling.knative.dev/class: kpa.autoscaling.knative.dev
    autoscaling.knative.dev/metric: concurrency
    autoscaling.knative.dev/target: "10"  # Target 10 concurrent requests
spec:
  predictor:
    minReplicas: 1
    maxReplicas: 100
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://models/fraud-detector/v1

Common Mistakes

Mistake	Problem	Solution
No latency budget	Slow predictions	Set SLOs, profile regularly
Ignoring cold starts	First request slow	Pre-warm, min replicas > 0
No batching	Inefficient GPU usage	Batch requests for throughput
No fallback	Service failure = outage	Default predictions, circuit breakers
Wrong serving pattern	Over-engineered or too slow	Match pattern to use case
No model versioning	Can’t rollback	Version every deployment

Quiz

Test your understanding:

1. When would you choose batch serving over real-time serving?

Answer: Choose batch serving when:

• Latency isn’t critical: Recommendations can be hours old
• High volume: Pre-compute for all users is efficient
• Complex models: Slow models can run offline
• Cost optimization: Batch is cheaper than real-time

Real-time is needed when:

• Immediate response required (fraud detection)
• Input is dynamic (can’t pre-compute all possibilities)
• Freshness matters (current context affects prediction)

2. What's the difference between canary and shadow deployments?

Answer:

• Canary: New model serves a percentage of real traffic. Users receive predictions from the new model. Risk: Bad predictions affect some users.

• Shadow: New model receives real traffic but responses are logged, not served. Users always get predictions from the current model. Risk: None to users, but can’t measure user impact.

Use shadow first (validate technical metrics), then canary (validate business metrics).

3. Why is model optimization important for serving?

Answer: Serving costs dominate ML costs:

• Training: Once or occasionally
• Serving: Continuous, at scale

Optimization reduces:

• Latency: Better user experience
• Cost: Fewer resources per prediction
• Scale: More predictions per dollar
• Memory: More models per server

A 2x faster model can mean 50% cost savings—significant at scale.

4. What causes cold start problems and how do you mitigate them?

Answer: Cold start occurs when:

• Model loads into memory for first request
• Pods scale from zero
• New replicas initialize

Mitigation strategies:

• minReplicas > 0: Always have warm replicas
• Pre-warm: Send synthetic requests before traffic
• Smaller models: Faster to load
• Model caching: Keep models in shared storage
• Predictive scaling: Scale before traffic spike

Hands-On Exercise: Deploy a Model with KServe

Deploy a model and test different serving patterns:

Setup

# Requires: Kubernetes cluster with KServe installed
# For local testing, use kind with KServe quickstart

# Create namespace
kubectl create namespace ml-serving

# Create S3-compatible storage (MinIO for local)
kubectl apply -f https://raw.githubusercontent.com/kserve/kserve/master/hack/quick_install.sh

Step 1: Train and Save Model

train_and_export.py

import joblib
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

# Train
X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

model = RandomForestClassifier(n_estimators=100)
model.fit(X_train, y_train)

print(f"Accuracy: {model.score(X_test, y_test):.4f}")

# Save
joblib.dump(model, "model.joblib")
print("Model saved to model.joblib")

Step 2: Create InferenceService

inference-service.yaml

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: iris-classifier
  namespace: ml-serving
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: gs://your-bucket/iris-model  # Or s3://, pvc://, etc.
      resources:
        requests:
          cpu: "100m"
          memory: "256Mi"
        limits:
          cpu: "500m"
          memory: "512Mi"
    minReplicas: 1
    maxReplicas: 5

Step 3: Deploy and Test

# Deploy
kubectl apply -f inference-service.yaml

# Wait for ready
kubectl wait --for=condition=Ready inferenceservice/iris-classifier -n ml-serving --timeout=300s

# Get URL
SERVICE_URL=$(kubectl get inferenceservice iris-classifier -n ml-serving -o jsonpath='{.status.url}')
echo "Service URL: $SERVICE_URL"

# Test prediction
curl -X POST "${SERVICE_URL}/v1/models/iris-classifier:predict" \
  -H "Content-Type: application/json" \
  -d '{
    "instances": [
      [5.1, 3.5, 1.4, 0.2],
      [6.2, 3.4, 5.4, 2.3]
    ]
  }'

Step 4: Set Up Canary Deployment

canary-deployment.yaml

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: iris-classifier
  namespace: ml-serving
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: gs://your-bucket/iris-model-v1
    canaryTrafficPercent: 20
  canary:
    predictor:
      model:
        modelFormat:
          name: sklearn
        storageUri: gs://your-bucket/iris-model-v2

Step 5: Monitor and Promote

# Check traffic split
kubectl get inferenceservice iris-classifier -n ml-serving -o yaml | grep -A 5 "traffic"

# If canary performs well, promote (set canary to 100%)
kubectl patch inferenceservice iris-classifier -n ml-serving --type='merge' -p '
spec:
  predictor:
    canaryTrafficPercent: 100
'

# Or rollback (set canary to 0%)
kubectl patch inferenceservice iris-classifier -n ml-serving --type='merge' -p '
spec:
  predictor:
    canaryTrafficPercent: 0
'

Success Criteria

You’ve completed this exercise when you can:

• Deploy a model as an InferenceService
• Send prediction requests
• Set up canary deployment with traffic splitting
• Promote or rollback canary
• View autoscaling behavior

Key Takeaways

1. Choose the right serving pattern: Online, batch, streaming, or edge—match to use case
2. KServe provides production-grade serving: Autoscaling, canary, multi-framework
3. Optimize models for serving: Quantization, compilation, batching
4. Use canary deployments: Gradual rollout reduces risk
5. Plan for cold starts: Pre-warm, min replicas, fast loading

Further Reading

• KServe Documentation — Kubernetes model serving
• Seldon Core — Enterprise ML serving
• ONNX Runtime — Cross-platform inference
• TensorRT — NVIDIA GPU optimization

Summary

Model serving bridges the gap between training and production. The right serving pattern (online, batch, streaming, edge) depends on latency requirements and use case. KServe provides a Kubernetes-native solution with autoscaling, canary deployments, and multi-framework support. Optimization techniques (quantization, compilation) reduce latency and cost. Gradual rollouts (canary, shadow) reduce deployment risk.

---

Next Module

Continue to Module 5.5: Model Monitoring & Observability to learn how to detect model degradation before it impacts users.