Deploy ML Model Serving

See Extended Examples for complete configuration files and templates.

ML → prod. Scalable serving, monitoring, A/B.

Use When

• Trained models → prod real-time inference
• REST/gRPC APIs → predictions
• Autoscale → variable load
• A/B tests → model vers
• Batch → real-time migrate
• Low-latency prediction svcs
• Multi-ver mgmt prod

In

• Required: Registered model (MLflow Model Registry) or trained artifact
• Required: K8s or container orchestration
• Required: Serving framework (MLflow, BentoML, Seldon Core, TorchServe)
• Optional: GPU → deep learning
• Optional: Monitoring (Prometheus, Grafana)
• Optional: LB + ingress

Do

Step 1: MLflow Models Serving

Built-in → quick sklearn/PyTorch/TF.

# Serve model locally for testing
mlflow models serve \
  --model-uri models:/customer-churn-classifier/Production \
  --port 5001 \
  --host 0.0.0.0

# Test endpoint
curl -X POST http://localhost:5001/invocations \
  -H 'Content-Type: application/json' \
  -d '{
    "dataframe_records": [
      {"feature1": 1.0, "feature2": 2.0, "feature3": 3.0}
    ]
  }'

Docker deploy:

# Dockerfile.mlflow-serving
FROM python:3.9-slim

# Install MLflow and dependencies
RUN pip install mlflow boto3 scikit-learn

# Set environment variables
ENV MLFLOW_TRACKING_URI=http://mlflow-server:5000
# ... (see EXAMPLES.md for complete implementation)

Docker Compose:

# docker-compose.mlflow-serving.yml
version: '3.8'

services:
  model-server:
    build:
      context: .
      dockerfile: Dockerfile.mlflow-serving
# ... (see EXAMPLES.md for complete implementation)

Test:

# test_mlflow_serving.py
import requests
import json

def test_prediction():
    url = "http://localhost:8080/invocations"

    # Prepare input data
# ... (see EXAMPLES.md for complete implementation)

→ Server starts, HTTP POST OK, JSON predictions, Docker runs clean.

If err: Model URI valid (mlflow models list), tracking server reachable, deps in container, port free (netstat -tulpn | grep 8080), flavor compat, docker logs <container-id>.

Step 2: BentoML → prod scale

Advanced serving, better perf.

# bentoml_service.py
import bentoml
from bentoml.io import JSON, NumpyNdarray
import numpy as np
import pandas as pd

# Load model from MLflow
import mlflow
# ... (see EXAMPLES.md for complete implementation)

Build + containerize:

# Build Bento
bentoml build

# Containerize
bentoml containerize customer_churn_classifier:latest \
  --image-tag customer-churn:v1.0

# Run container
docker run -p 3000:3000 customer-churn:v1.0

BentoML config:

# bentofile.yaml
service: "bentoml_service:ChurnPredictionService"
include:
  - "bentoml_service.py"
  - "preprocessing.py"
python:
  packages:
    - scikit-learn==1.0.2
    - pandas==1.4.0
    - numpy==1.22.0
    - mlflow==2.0.1
docker:
  distro: debian
  python_version: "3.9"
  cuda_version: null  # Set to "11.6" for GPU support

K8s deploy:

# k8s/deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: churn-prediction
  labels:
    app: churn-prediction
spec:
# ... (see EXAMPLES.md for complete implementation)

Deploy → K8s:

# Apply Kubernetes manifests
kubectl apply -f k8s/deployment.yaml

# Check deployment status
kubectl get deployments
kubectl get pods
kubectl get services

# Test endpoint
EXTERNAL_IP=$(kubectl get svc churn-prediction-service -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
curl -X POST http://$EXTERNAL_IP/predict \
  -H 'Content-Type: application/json' \
  -d '{"instances": [{"tenure": 12, "monthly_charges": 70.35}]}'

→ Bento builds, container serves, K8s 3 replicas, LB external EP, health OK.

If err: bentoml --version, model in store (bentoml models list), Docker running, K8s access (kubectl cluster-info), resource limits, pod logs (kubectl logs <pod-name>), svc selector matches labels.

Step 3: Seldon Core → advanced

Multi-model serving, A/B, explainability.

# seldon_wrapper.py
import logging
from typing import Dict, List, Union
import numpy as np
import mlflow

logger = logging.getLogger(__name__)

# ... (see EXAMPLES.md for complete implementation)

Seldon deploy config:

# seldon-deployment.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier
  namespace: seldon
spec:
  name: churn-classifier
# ... (see EXAMPLES.md for complete implementation)

A/B test:

# seldon-ab-test.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier-ab
spec:
  name: churn-classifier-ab
  predictors:
# ... (see EXAMPLES.md for complete implementation)

Deploy:

# Install Seldon Core operator
kubectl create namespace seldon-system
helm install seldon-core seldon-core-operator \
  --repo https://storage.googleapis.com/seldon-charts \
  --namespace seldon-system \
  --set usageMetrics.enabled=true

# Create namespace for models
# ... (see EXAMPLES.md for complete implementation)

→ Seldon operator OK, pods created, REST EP responds, A/B splits traffic, analytics records.

If err: Operator (kubectl get pods -n seldon-system), SeldonDeployment status (kubectl describe seldondeployment), image registry access, model URI resolution, RBAC, model container logs.

Step 4: Monitoring + observability

Comprehensive metrics.

# monitoring.py
from prometheus_client import Counter, Histogram, Gauge, start_http_server
import time
import logging

logger = logging.getLogger(__name__)

# Prometheus metrics
# ... (see EXAMPLES.md for complete implementation)

Prometheus config:

# prometheus-config.yaml
global:
  scrape_interval: 15s
  evaluation_interval: 15s

scrape_configs:
  - job_name: 'model-serving'
    kubernetes_sd_configs:
# ... (see EXAMPLES.md for complete implementation)

Grafana JSON:

{
  "dashboard": {
    "title": "ML Model Serving Metrics",
    "panels": [
      {
        "title": "Predictions Per Second",
        "targets": [
          {
# ... (see EXAMPLES.md for complete implementation)

→ Prometheus scrapes OK, Grafana shows throughput + latency + err rates + active reqs real-time.

If err: Scrape targets UP (http://prometheus:9090/targets), metrics EP (curl http://model-pod:8000/metrics), K8s svc discovery, datasource, firewall port.

Step 5: Autoscaling

HPA by req load.

# hpa.yaml
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: churn-prediction-hpa
  namespace: seldon
spec:
  scaleTargetRef:
# ... (see EXAMPLES.md for complete implementation)

Apply:

# Enable metrics server (if not already installed)
kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

# Apply HPA
kubectl apply -f hpa.yaml

# Check HPA status
kubectl get hpa -n seldon
kubectl describe hpa churn-prediction-hpa -n seldon

# Load test to trigger scaling
kubectl run -it --rm load-generator --image=busybox --restart=Never -- /bin/sh -c "while sleep 0.01; do wget -q -O- http://churn-prediction-service/predict; done"

# Watch scaling
kubectl get hpa -n seldon --watch

→ HPA monitors CPU/mem/custom, scales up on load, down after stabilize, min/max respected.

If err: metrics-server (kubectl get deployment metrics-server -n kube-system), pod resource reqs defined, custom metrics available, RBAC, stabilize windows.

Step 6: Canary deploy

Traffic shift.

# canary-deployment.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier-canary
spec:
  name: churn-classifier-canary
  predictors:
# ... (see EXAMPLES.md for complete implementation)

Gradual rollout:

# canary_rollout.py
import time
import subprocess
import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# ... (see EXAMPLES.md for complete implementation)

→ Canary 0%, gradual shift, health OK each stage, rollback if degrade, full rollout after all pass.

If err: Multi predictors, traffic sums 100%, canary image pullable, Prometheus metrics for health, rollback logic, both ver logs.

Check

• Server responds → prediction req
• REST/gRPC EPs OK + docs
• Docker containers build + run
• K8s creates expected replicas
• LB → external EP
• Liveness/readiness pass
• Prometheus scraped
• Grafana real-time
• Autoscale on load
• A/B splits correctly
• Canary gradual rollout
• Rollback works

Traps

• Cold start: First req slow → readiness probe delay, cache model
• Mem leaks: Accumulate → monitor, periodic restart, profile
• Dep conflicts: → exact pinned vers, test Docker pre-deploy
• Resource limits low: OOM/throttle → profile, set by load test
• No health checks: K8s routes to unhealthy → liveness/readiness probes
• No rollback: Bad deploy → canary, keep prev ver
• Ignore latency: Only accuracy → bench, optimize, batch
• Single replica: No HA → min 2, anti-affinity
• No monitoring: Until complaints → metrics day 1
• GPU unused: → CUDA visible devices, K8s alloc

→

• register-ml-model — register before deploy
• run-ab-test-models — A/B ver testing
• deploy-to-kubernetes — K8s patterns
• monitor-ml-model-performance — drift + degrade
• orchestrate-ml-pipeline — auto retrain + deploy