Alejandro Saucedo - The State of Production Machine Learning in 2023 | PyData Global 2023

Machine learning model lifecycle doesn’t end at training - it begins when deployed to production and requires ongoing monitoring, maintenance and governance

Organizations are moving from model-centric to data-centric approaches, with increased focus on data quality, versioning, and metadata management

MLOps requires cross-functional teams including ML engineers, data scientists, DevOps, and domain experts - the “unicorn” data scientist model doesn’t scale

Security considerations must be integrated at every stage of the ML pipeline, not just deployment - includes data poisoning, model artifacts, access control

Standardization and best practices are emerging around ML governance, including responsible AI frameworks, monitoring, and compliance requirements

Organizational structures are evolving with dedicated MLOps teams and platform engineering capabilities to support ML at scale

Monitoring needs to cover both traditional metrics (CPU, RAM) and ML-specific concerns like data drift, model performance degradation

The impact of ML systems remains fundamentally human-focused regardless of technical complexity - ethical considerations can’t fall on individual practitioners

Organizations should avoid adopting too much tooling too early - start with basic foundations and scale complexity as needed

LLMs and agent architectures introduce new operational challenges around compute resources, monitoring and security that differ from traditional ML systems