Data & ML Engineering · 300-level

Data & Machine Learning Engineering

A staff-level tour of the data and ML engineering stack — warehouse modeling, batch and streaming pipelines, data quality, feature stores, training and evaluation, deployment, drift monitoring, and production RAG — taught through mechanisms, real tools, and the dollar-and-latency consequence of getting each one wrong.

DML-3013 creditselectiveno prerequisites
What's inside

Sections & lessons

01

Data modeling & warehouses

  • OLTP vs OLAP: why analytics doesn't belong on your production databaseconcept40 min
  • Dimensional modeling: star vs snowflake and the discipline of grainconcept40 min
  • Columnar storage and partitioning: how the warehouse scans less to cost lessconcept35 min
  • Slowly-changing dimensions: keeping history without corrupting the presentconcept35 min
02

Batch & streaming pipelines

  • ETL vs ELT: where the transform runs and who pays for itconcept40 min
  • Spark at scale: partitions, shuffles, and the skew that stalls a jobconcept40 min
  • Streaming with Kafka: partitions, ordering, and delivery semanticsconcept35 min
  • Orchestration and idempotent backfills: Airflow, Dagster, and replay-safetyconcept35 min
03

Data quality & governance

  • Data contracts: schemas as enforceable agreementsconcept40 min
  • Testing data: dbt tests and Great Expectationsdemo40 min
  • Lineage and cataloging: knowing what breaks three hops downstreamconcept35 min
  • PII handling and governance: classification, masking, and accessconcept35 min
04

Feature engineering & feature stores

  • Data leakage: the bug that inflates every offline metricconcept40 min
  • Train/serve skew: when production features drift from trainingconcept40 min
  • Feature stores: offline/online parity as an architectureconcept35 min
  • Point-in-time correctness: as-of joins that don't leak the futureconcept35 min
05

Model training & evaluation

  • Splitting data honestly: temporal and grouped splitsconcept40 min
  • Choosing the metric that matches the decisionconcept40 min
  • Overfitting, baselines, and cross-validationconcept35 min
  • Imbalanced data: when accuracy liesconcept35 min
06

MLOps & deployment

  • Serving patterns: batch, online, and streaming inferenceconcept40 min
  • Versioning and reproducibility: pinning model, data, and code togetherconcept40 min
  • CI/CD for ML: testing a model before it shipsconcept35 min
  • Shadow and canary deploys: validating in production without betting the businessdemo40 min
07

Monitoring & drift

  • Data drift vs concept drift: two failures that look alike on a dashboardconcept40 min
  • Monitoring models when labels arrive lateconcept35 min
  • Retraining triggers: deciding when to retrain and proving it helpedconcept35 min
08

LLMs & RAG in production

  • Embeddings and vector databases: retrieval as nearest-neighbor searchconcept40 min
  • Retrieval quality and RAG evals: measuring what the retriever actually returnsconcept40 min
  • Guardrails, cost, and latency in production RAGconcept35 min

This module ends in a gate you can fail.

That's what makes passing it mean something. Take the DSAT, get placed, and start earning.