Podcast Atlas

EP 7: Data Science & MLOps

source Episode summary Updated 2026-08-18 Tags: Podcast, Data-Science, Mlops, Ml-Engineering

Summary

This Data Science With Sam episode has Sam interview Aaron Blythe, a Google Cloud customer engineer with application-development, DevOps, and machine-learning interests, about how MLOps, Machine Learning Engineering, and Data Engineering For Data Science fit around data science. The conversation connects production ML to DevOps CALMS, ML CI/CD, Production ML Feedback Loops, Integrated ML Teams, and Data Scientist MLOps Fluency. Its core synthesis is that data scientists do not need to own every deployment and operations task, but credible industry ML depends on shared language, automation, measurement, feedback, data infrastructure, and cross-trained teams.

Key Claims

  • Aaron Blythe describes himself as a Google Cloud customer engineer who moved from application development in healthcare into DevOps and then machine-learning work.
  • The episode frames data science as a hybrid of math, programming, curiosity, research reading, and practical business application.
  • Aaron says important data-science traits include curiosity, open-mindedness, perseverance, and patience, not only statistics, linear algebra, R, or Python.
  • The source distinguishes scientists, who form hypotheses and reusable findings, from engineers, who implement, keep systems moving, and align work with organizational goals.
  • Data Engineering For Data Science is presented as the foundation that gets data into the right place so data scientists are not trapped in local CSV wrangling.
  • Aaron uses data warehouses such as BigQuery as examples of letting analysis happen close to the data rather than repeatedly copying data into local files.
  • Machine Learning Engineering is framed as the work of putting models somewhere useful, often behind an API, and connecting production behavior back to model improvement.
  • Production ML Feedback Loops matter because user behavior, missing features, and new knowledge should return to the data scientist after deployment.
  • Sam ties production ML to concept drift and data drift, arguing that no model is foolproof and that deployed models need continuing improvement.
  • DevOps CALMS enters through Aaron’s explanation of DevOps as culture, automation, lean or learning, measurement, and sharing.
  • MLOps borrows DevOps ideas for model deployment: automate delivery, measure behavior, share responsibility, and close the feedback loop.
  • ML CI/CD is described as a target state for MLOps, but Aaron says he has not seen a company that has fully perfected it.
  • The episode treats MLOps as partly experimental because teams are still discovering which production-ML practices work best.
  • Data Scientist MLOps Fluency is the practical role boundary: data scientists should understand MLOps well enough to collaborate, but they do not need to perform every hands-on operations task.
  • Integrated ML Teams are the preferred organizational pattern, combining data engineers, data scientists, and ML engineers around business functions such as recommendation systems.

Key Quotes

“short answer is yes” - Aaron’s answer on whether data scientists should understand MLOps.

“if you write it, you run it” - DevOps ownership principle the episode connects to MLOps team design.

“some science” - Aaron’s description of current MLOps practice as still partly experimental.

Connections

Contradictions

  • No direct contradiction found.
  • The source qualifies broad data-science and AI-career optimism by showing that useful ML is not only model building: production value depends on data engineering, deployment, APIs, CI/CD, measurement, drift response, team ownership, and cross-role collaboration.