MLOps Fundamentals: Running ML in Production

Fundamentals

Principles

• Reproducibility and traceability
• Automation and observability
• Testing, governance, and security by design

Lifecycle

• Data management and feature pipelines
• Experimentation and model training
• Validation, deployment, and monitoring
• Feedback loops and continuous improvement

Architecture Patterns

• Batch vs real-time inference
• Online/offline feature stores
• Shadow, canary, and blue/green deploys

How the Algorithm Works

Release

• Build → validate → deploy via CI/CD
• Shadow/canary → blue/green strategies

Retrain

• Data triggers and schedules
• Automated evaluation and approvals

Rollback

• Guardrails, SLO breaches → revert
• Versioned artifacts and configs

Real-World Applications

KPIs and Constraints

Latency, accuracy, cost, fairness, and privacy drive MLOps designs and rollouts.

Best Practices

Checklist

• Version data, code, and models
• Automate CI/CD and retraining
• Monitor data, model, and system KPIs
• Secure secrets and PII
• Govern with approvals and audits

Anti-Patterns

• Notebook-only deployments
• No monitoring or rollback plan
• Leaking test sets into training
• Untracked experiments and artifacts
• Ignoring fairness and compliance

Worked Examples

Batch Pipeline

# Orchestrate ETL → train → validate → register → deploy

Real-Time Inference

# Feature store + online model serving with canary

Implementation

Pipelines

• Orchestrate with Airflow/Prefect
• Data validation and drift checks

Model Registry

• Version models and manage stages
• Approval flows and lineage

Observability

• Metrics, logs, traces; model-specific monitors
• Dashboards and alerts

Security

• Secrets, PII controls, and SBOMs
• Network policies and isolation

The Math Behind It

SLAs and Error Budgets

Define SLOs (latency, uptime, accuracy) and track error budgets to guide releases and rollbacks.

Queueing Models

Use M/M/k approximations to plan capacity and meet P95 latency at target QPS.