Wiki
MLOps
Podcast-grounded reference page for MLOps as the operating discipline for production machine learning systems.
Related Wiki Pages
MLOps is the operating discipline for machine learning systems after they leave experimentation. It starts with reproducible training and model artifacts. It continues through deployment and serving. It also covers monitoring, retraining, governance, and ownership.
In Building Production ML Platforms, Simon Stiebellehner defines MLOps through people, workflows, and technology around 4:42. In MLOps at Scale, Raphaël Hoogvliets connects that operating discipline to CI and repository structure around 39:06-56:50. He also covers reproducibility, serving, monitoring, and package management.
Use DataOps for the separate discipline around data pipelines and analytical delivery. Use MLOps and DataOps and MLOps vs DataOps for the boundary between model operations and data operations. The MLOps side adds model artifacts and experiment capture. It also adds drift, retraining, and model governance.
Related MLOps Areas
MLOps connects most directly to these wiki pages:
- ML Platforms
- Machine Learning System Design
- Model Registry
- Model Monitoring
- Experiment Tracking
- Reproducibility
- Production
Core podcast interviews include:
- MLOps at Scale with Raphaël Hoogvliets
- Building Production ML Platforms with Simon Stiebellehner
- Pragmatic MLOps with Maria Vechtomova
- Mastering MLOps with Theofilos Papapanagiotou
- MLOps Architect Guide with Danny Leybzon
- Human-Centered MLOps with Lina Weichbrodt
- MLOps in Finance and Lean MLOps for Startups with Nemanja Radojkovic
Common Definition
Across the archive, MLOps means the repeatable path from model development to a maintained production system. In MLOps at Scale, Raphaël covers adoption by product teams around 23:01-36:55. Around 39:06-56:50, he moves into CI, repository structure, and parameterization. He also covers reproducibility and monitoring. Raphaël also covers package registries and containers.
In Building Production ML Platforms, Simon frames MLOps through people, workflows, and technology around 4:42. Around 28:20-54:15, he walks through experiment tracking and model registries. He then covers batch and online serving, orchestration, and metadata. The same sequence ties those pieces to lineage and governance.
The common thread is ownership after training. MLOps doesn’t stop when a notebook produces a good metric. It asks whether another person can reproduce a run and approve an artifact. It also asks whether the team can deploy and monitor the model. The team also needs rollback, retraining, and retirement decisions.
Theofilos makes that lifecycle boundary explicit in Mastering MLOps around 7:28-15:29. He contrasts DevOps with model lifecycle, drift, and fairness. He also covers monitoring and retraining triggers.
Guest Tradeoffs
Guests differ on how much platform to build and how early to build it. In Lean MLOps for Startups, Nemanja Radojkovic emphasizes minimal operational discipline for small teams around 7:54-21:35. Small teams can use SaaS and managed services when those tools help them launch, but they still need to watch vendor lock-in and operational overhead. In Building Production ML Platforms, Simon makes the platform layer more important when multiple teams repeat the same training and deployment work. Around 16:52-20:04 and 47:08-49:19, he also connects that platform need to registries and monitoring.
Guests also differ on tooling emphasis. In Pragmatic MLOps, Maria Vechtomova stresses standard engineering primitives around 16:27-20:49. She names Git, CI/CD, and registries. She also names Kubernetes, reusable repositories, and monitoring.
In Human-Centered MLOps, Lina Weichbrodt shifts the emphasis toward business cases and stakeholder buy-in around 4:50-29:23. She also covers monitoring and incident response. Both views agree that tools aren’t the strategy. Maria starts from standardization and reuse, while Lina starts from the people who need to trust and debug the model.
Model Lifecycle
MLOps begins when a model must become a maintained system. Teams need tracked experiments and model artifacts. They also need approval and deployment.
Teams need serving and monitoring after deployment. Once the model runs, teams need feedback and retraining decisions.
The Model Registry and Experiment Tracking pages cover the handoff from training to production. Simon describes that handoff in Building Production ML Platforms around 29:41-31:51. Experiment tracking leads into registries and serving, then into batch inference and orchestration.
MLOps at Scale adds the reproducibility side of the same lifecycle around 42:31-53:08. Raphaël covers data versioning and traceability. He also covers experiment capture and model registries. He also covers serving, monitoring, and dependency management. That distinction matters because a daily batch scoring job, a low-latency API, and a managed endpoint have different failure modes and rollback paths.
Monitoring and Feedback
MLOps monitoring covers model behavior, service behavior, and the data around the model. In MLOps Architect Guide, Danny Leybzon connects model monitoring to upstream ETL, data pipelines, and observability around 25:04-34:25.
A model can degrade because features shifted or labels arrived late. It can also degrade because a schema changed or the serving path broke. Lina adds the product operations side in Human-Centered MLOps around 24:34-49:28. She covers service levels and post-mortems. She also covers live test sets and user bug reports.
Her monitoring discussion includes input distribution checks, feature drift, and logging. She also covers feature stores and reproducibility.
Use Model Monitoring for the model-specific layer. Use DataOps and Data Quality and Observability when the root cause sits in the data pipeline rather than the model artifact. Theofilos marks that overlap around 50:35 in Mastering MLOps, where the discussion turns to the continuation and divergence between DataOps and MLOps.
Governance and Risk
MLOps becomes stricter when models affect regulated decisions in finance and healthcare. The same concern applies to fraud, risk, and customer-facing workflows. In MLOps in Finance, Nemanja ties deployment to CI/CD, monitoring, and governance around 14:57-23:39. He also covers validation and risk controls.
He later describes a minimal finance stack around 31:57-35:57 with dev, test, and production environments. In the same discussion, he covers monitoring, model versioning, and simple interim registries. Those requirements make approval history, lineage, model versioning, and rollback paths part of the system design.
This is where Production, Reproducibility, and Governance connect to MLOps. A team needs to explain which model produced which output. It also needs the data, approval path, and monitoring path. Simon covers the same governance trail around 42:48-45:50 in Building Production ML Platforms, where metadata and lineage influence the platform design. He also covers artifact logging, tracking, and GDPR concerns.
Related Pages
These pages cover adjacent MLOps concepts: