All stacks · AI & MLOps
Feature stores, model pipelines, and ML infrastructure.
MLOpsThe MLOps landscape explained through a data engineering lens: what feature stores actually solve, why training pipelines are just DAGs, and the skills that transfer.
RedisWe trained on warehouse snapshots and served from Redis. TTL eviction and a stampede later, the model looked broken. Online features are a cache, not a fact table.
MLOps is where data engineering meets model lifecycle: feature pipelines, training data versioning, deployment, and the monitoring that notices drift before customers do. Most "ML platform" pain is data platform pain wearing a different badge.
The guides here approach MLOps from the pipeline side — feature stores and when a well-modeled warehouse beats one, reproducible training data, batch versus online serving, and monitoring that watches data quality as closely as model metrics.
Only when online inference needs the same features as training at low latency, or many teams reuse features. For batch scoring, versioned warehouse tables built by dbt are simpler and usually sufficient. Adopt the problem before the tool.
When features computed at serving time differ from how they were computed for training — different code paths, different timing, leaked future data. It quietly destroys model performance. The cure is computing features once, in one pipeline, consumed by both sides.
Code is only one of three moving parts — data and models change too, often without a deploy. That means versioning datasets, tracking experiments, and monitoring statistical behavior in production, not just uptime.
Input data quality first (nulls, ranges, schema, volume), then prediction distribution drift, then delayed ground-truth metrics where labels arrive late. Most "model degraded" incidents are upstream data incidents caught late.
MLflow for tracking and registry (it is everywhere), plus your platform's serving story — SageMaker, Vertex, or Databricks. Orchestration you already know from data work transfers directly; a model pipeline is still a DAG.
One email a week with our latest tutorials. No spam.
Newsletter signup is not live yet. Use the contact form if you want to be notified.