Model Registry Engineering: Artifact Tracking, Metadata, Lineage, Promotion Workflows, and Governance for ML Models - Tapa blanda

Sinopsis

Model Registry Engineering is a systems-oriented guide to the infrastructure that determines whether your team can answer, months after a deployment, exactly what model was running and why.

A model registry is not a file server with version numbers. It is the system of record connecting every production model back to the training run, data version, code commit, and approval that produced it — and forward to every environment it was deployed to.

Thirteen chapters cover the full registry stack:

• Registry foundations: what constitutes a complete artifact, immutable versioning, and lifecycle stages that mean something
• Artifact tracking: weights, configs, tokenizers, and checksum verification at load time
• Metadata: metrics with their evaluation protocol, dataset versions, framework and hardware context
• Lineage: training runs, data versions, code commits, and dependency locks — queryable in both directions
• Promotion workflows: candidate gates, staging parity, registry-driven deployment, and evidence-backed approval
• Governance: access control, risk-scaled review, retention policy, and complete audit trails
• Deployment integration: release pipelines, serving reconciliation, and rollback as a first-class operation
• Registry quality: completeness, consistency, staleness detection, and write-time validation
• Failure modes: orphan artifacts, wrong lineage, untracked copies, and stage drift
• Registry APIs and automation: required fields in signatures, idempotent writes, and event streams
• Multi-team registries: namespace design, shared base models, cost attribution, and visibility defaults
• Cost and scale: storage growth, deduplication, retrieval patterns, and archival economics
• Production operations: backup and restore, safe migration, correctness monitoring, and policy change management

Every section pairs the concept with the trade-off that makes it a real engineering decision, the pitfall teams most often hit, and a short set of practical checks. Each chapter closes with a worked scenario showing how these decisions interact under real pressure.

Three appendices cover architectural comparisons (centralized versus federated, content-addressed versus sequential versioning, registry-pull versus image-baked deployment), a four-level registry maturity model, and adoption sequencing for teams starting from scratch.

Written for ML platform engineers, MLOps practitioners, and technical leads responsible for what actually runs in production.

"Sinopsis" puede pertenecer a otra edición de este libro.