Ml Model Deployment

ML Model Deployment

Deploy models to production with serving infrastructure that is reliable, observable, and safely updatable. Deploying a model is not the finish line — it's the beginning of the operational lifecycle.

Key Difference from Software Deployment

Software deployment is primarily concerned with "does the code work?" Model deployment adds: "does the model still work well on real-world data that changes over time?" Models degrade silently — they keep returning predictions, just increasingly wrong ones.

Workflow

Step 1: Package the Model

Before deployment, the model must be a self-contained, versioned artifact:

• Serialize the model: Use framework-native formats (joblib/pickle for sklearn, SavedModel for TF, TorchScript for PyTorch, ONNX for cross-framework)
• Pin dependencies: Exact versions of libraries used during training
• Include preprocessing: The serving pipeline must apply the same feature transformations as training. Feature skew (different preprocessing in training vs serving) is the #1 cause of silent model failure
• Version it: Tag with model registry version, git commit, training data version, and training metrics

Step 2: Choose the Serving Pattern

For most teams starting out: REST API with FastAPI is the right default. Optimize for lower latency only when measured need justifies complexity.

Step 3: Build the Serving Application

The serving app needs:

Request → Validate input → Preprocess → Predict → Postprocess → Response

Key requirements:

• Input validation: Reject malformed requests before they reach the model. With FastAPI this is a Pydantic request model on the predict endpoint — declare the feature schema with types and ranges, and return 422 on a bad payload (see references/serving-patterns.md). Under a high-throughput requirement this is the first thing people cut, and it is the wrong cut. Trimming pandas, fixing column order, and batching are the latency wins; the request schema is not one. An unvalidated fraud endpoint scores garbage features silently instead of rejecting them — a wrong prediction costs more than the microseconds the check takes.
• Feature preprocessing: Apply the same transformations used in training
• Error handling: Model errors should return 500 with diagnostic info, not crash the service
• Logging: Log every prediction request/response for monitoring and debugging
• Health check: Endpoint that verifies the model is loaded and responsive

See references/serving-patterns.md for implementation templates.

Step 4: Deploy Safely

Models should be deployed with the same caution as any production service — more, given their non-deterministic nature:

Shadow mode: Deploy the new model alongside the current one. Both receive traffic, but only the current model's predictions are used. Compare predictions to evaluate the new model on real data.

Canary deployment: Route a small percentage of traffic (5-10%) to the new model. Monitor metrics. Gradually increase if performance is acceptable.

A/B testing: Split traffic between model versions to measure business impact (conversion rate, click-through, etc.), not just ML metrics.

Blue-green: Deploy new model to a separate environment, switch all traffic at once after validation. Fastest rollback.

Step 5: Monitor in Production

After deployment, monitor continuously. See references/serving-patterns.md for implementation details.

Model performance monitoring:

• Track prediction accuracy against ground truth (when available)
• Monitor prediction distribution (are predictions shifting?)
• Track business KPIs that the model influences

Data drift detection:

• Compare input feature distributions against the training data baseline
• Use statistical tests: Population Stability Index (PSI), Kolmogorov-Smirnov (KS) test, Jensen-Shannon divergence
• Alert when drift exceeds thresholds

System monitoring:

• Prediction latency (P50, P95, P99)
• Request throughput and error rates
• Resource utilization (CPU, memory, GPU)
• Model loading time

Retraining triggers:

• Accuracy drops below threshold
• Data drift exceeds threshold
• Scheduled retraining (weekly/monthly depending on domain)
• New labeled data becomes available

Step 6: Plan Rollback

Every model deployment needs a rollback plan:

• Previous model version tagged and ready in the registry
• One-command rollback through CI/CD or model registry stage transition
• Automated rollback on monitoring alert (if confidence is high)

Principles Applied

• KISS: Start with REST API + batch monitoring. Add real-time drift detection when scale justifies it.
• YAGNI: Don't build an A/B testing framework for your first model. Shadow mode is enough.
• Defense in depth: Validate inputs, monitor outputs, detect drift, and have rollback ready.
• Functional Independence: Serving, monitoring, and retraining should be independent components, not monolithic.