Day 4 - Operations & Cost: Cost Optimization, Logging/Audit, Disaster Recovery

ML workflows cost continuously (data pipelines, training GPUs, always-on endpoints). Regulation demands "who did what when." Infrastructure fails. Today's three pillars: cost, audit, resilience.

Cost Optimization: Training vs Inference Split

Learning and serving have opposite economics.

Training Cost Cuts

• Managed Spot Training: 90% cheaper, checkpoint on interruption
• Right-sizing: No GPU if CPU suffices, distributed only if needed
• Early stopping in AMT: Discard hopeless trials early