MLOps: models are easiest
After three years in production MLOps, Emeka Boris argued “the model is the easiest part” — most work sits in data pipelines, monitoring, feature stores, retraining and infra costs, roughly 80% engineering vs 20% ML tweeted. The thread stresses maintenance and engineering effort as the real bottleneck for production reliability.
Emeka Boris Ama identifies as a machine‑learning engineer and data scientist on his GitHub profile [GitHub]. github.com The engineering‑first argument echoes long‑standing academic warnings: Google’s NeurIPS paper “Hidden Technical Debt in Machine Learning Systems” (2015) catalogued ML‑specific maintenance costs like entanglement and undeclared consumers that drive ongoing engineering effort. proceedings.neurips.cc Recent industry surveys quantify that gap — a Gartner‑summarized survey found roughly 54% of models move from pilot to production, and Gartner’s Feb 26, 2025 press release flagged “lack of AI‑ready data” as a top risk for AI projects. venturebeat.com Vendors and practitioner playbooks now prioritize pipelines and observability: Databricks publishes an MLOps Stacks pattern for CI/CD and reproducible asset bundles, while Google’s practitioner guide prescribes continuous training, monitoring, and metadata/feature management as operational essentials. docs.databricks.com
