Federated ML & AI Architecture

As of 2026-05-29

Federated ML/AI runs ML where it makes sense relative to the data — move compute to data (or share live) instead of reflexively copying everything into one place. In the BDC era (zero-copy sharing, M041; SAP↔Databricks, M034), federation is often the better architecture. Core question per use case: move data or move compute? Default to federation; copy only for concrete reasons (feature reuse, point-in-time snapshots). Data residency (GDPR M082, utilities/banking M099/M097) makes federation a requirement — centralising for ML is non-compliant; train/score where data lawfully resides, move only models/aggregates. Layers: governed products (M033) as feature source → feature store → training in Databricks → inference embedded back into SAP. Governance must hold across the seam (M077/M081/M034). Embed the round-trip (outputs reach decisions), don't strand ML in a lab. Honest caveat: federation adds design complexity — name the trade-offs (§0i). Cross-stack-architect frontier skill.