- MLOps applies DevOps discipline to ML: version control, CI/CD, testing, monitoring, and incident response.
- It covers the whole lifecycle: data, features, models, prompts, agents, and their retraining triggers.
- An ML service breaks when the code changes and also when the world changes, so drift monitoring is central.
- LLMOps is the LLM-specific slice; the two share most infrastructure.
- H2O MLOps combines model registry, deployment, monitoring, and drift detection.
What is MLOps?
MLOps is what stands between a notebook that works on the data scientist's laptop and a model that keeps making accurate predictions in production for years.
MLOps takes the ideas of DevOps (version control, continuous integration, continuous delivery, automated testing, monitoring, and on-call) and applies them to a system that has two extra moving parts DevOps doesn't: data and models. A software service breaks when the code changes. An ML service also breaks when the world changes: new user behavior, a shifted feature distribution, a labelling error upstream. MLOps is the discipline that catches those drifts before customers do.
Historically, MLOps grew out of the recognition that most ML projects never shipped, and the ones that did rarely survived their first year. The reasons were operational, not scientific: no model registry, no rollback, no monitoring, no clear ownership between data science and platform engineering. MLOps addresses all of these.
Today, MLOps also covers generative AI (LLMOps), agents, and the full machine learning lifecycle from feature engineering through retirement.
How does MLOps work?
A modern MLOps stack has six moving parts arranged as a closed loop: build → deploy → monitor → retrain.
- Feature store / feature engineering: features are built, versioned, and served consistently between training and inference. See the H2O Feature Store docs.
- Model registry / Model Hub: every trained model has a version, lineage, metrics, and approval state. See Model Hub and the H2O MLOps model registry docs.
- CI/CD for models: new candidates are tested against a fixed evaluation set before they get promoted; H2O Eval Studio automates this.
- Deployment: models are packaged (containers, MOJO scoring pipelines) and rolled out with canary/blue-green patterns. See H2O MLOps deployment docs.
- Monitoring and observability: live metrics on latency, throughput, prediction distribution, feature drift, and business KPIs; extended to AI observability for LLMs and agents.
- Retraining triggers: automated pipelines that detect model drift or evaluation regressions and kick off retraining or rollback.
MLOps vs. LLMOps vs. AIOps
MLOps
AIOps
Focus
Predictive ML models in production
LLM prompts, chains, agents, evals
Using AI to run IT operations
Key artifact
Trained model + features
Prompt + model + retrieval + eval
Runbook automation, anomaly detection
Change frequency
Weeks to months
Hours to days
Continuous
Primary metric
Prediction accuracy, business KPI
Groundedness, safety, cost per query
MTTR, alert precision
Benefits of MLOps
Models actually reach production. Standardized pipelines turn one-off notebooks into deployable artifacts.
Faster iteration. With CI/CD and evaluation, a new model can be tested and rolled out in hours instead of quarters.
Reliability. Monitoring catches drift, corrupt features, and label leakage before customers notice.
Governance and audit. Every model, feature, and prediction is versioned and traceable, a prerequisite for AI compliance in banking, healthcare, and government.
Challenges of MLOps
Ownership seams. Data science, platform, and application teams all touch the pipeline. Unclear ownership causes long outages.
Data drift is silent. A model can degrade for weeks before business KPIs move. Feature and prediction monitoring must be automatic.
Tool sprawl. A pieced-together stack of registry + serving + monitoring + drift is expensive to run. Integrated platforms simplify this.
MLOps use cases
- Financial services: credit, fraud, and AML models with strict audit and rollback requirements.
- Telecommunications: churn, next-best-action, and network anomaly models updated weekly.
- Healthcare: clinical decision-support models under FDA/HIPAA governance.
- Retail and e-commerce: recommendation and pricing models retraining on daily feedback.
- Public sector: models running in air-gapped, FedRAMP High environments.
MLOps in H2O AI
H2O MLOps brings model deployment, monitoring, drift detection, a model registry, and champion/challenger workflows together in one place. It works with H2O Driverless AI, H2O-3, custom PyTorch and TensorFlow models, and LLMs, across predictive AI, generative AI, and agents. H2O eScorer handles edge and external scoring. It runs on-prem and air-gapped for organizations that need it.