IA MLO study plan
IA MLO · Give each week a clear focus.
8 weeks · About 7 hours per week
MLOps engineering practice for machine learning lifecycle and deployment
Map the full ML lifecycle (problem framing, data ingestion, training, validation, deployment, monitoring) and write a one-paragraph description of each stage's inputs, outputs, and failure modes
Set up a local environment with a container runtime, a Python virtual environment, and a version control repository to host all hands-on work for the plan
Reproduce a minimal end-to-end pipeline: train a small scikit-learn model, serialize it, and serve predictions through a local HTTP endpoint
Define reproducibility controls (pinned dependencies, fixed random seeds, data and code versioning) and apply them to the pipeline from the previous goal
Your checkpoint
A public or local repository containing a working train-to-serve pipeline plus a one-page lifecycle map with reproducibility notes
MLOps engineering practice for machine learning lifecycle and deployment
Containerize the week-1 model server and run it with a pinned base image and a non-root user
Automate build, test, and image publication with a CI pipeline triggered on every commit
Write unit tests for preprocessing and prediction contracts, and a smoke test that starts the container and calls the endpoint
Add structured logging and an environment-variable configuration layer to the service
Your checkpoint
A green CI pipeline that builds, tests, and publishes a versioned container image on each push
MLOps engineering practice for machine learning lifecycle and deployment
Instrument training runs with an experiment-tracking tool to record parameters, metrics, and artifacts
Implement a data validation step that fails the pipeline on schema drift or out-of-range values
Compare at least three model configurations and register the best one with a stage label in a model registry
Document a promotion rule that decides when a model moves from staging to production
Your checkpoint
A tracked experiment report comparing three configurations plus a registry entry with a written promotion rule
MLOps engineering practice for machine learning lifecycle and deployment
Deploy the registered model behind a local orchestration layer with health checks and resource limits
Implement a canary or shadow release that routes a fraction of traffic to the new model version
Add automated rollback triggered by a failed health check or an error-rate threshold
Measure inference latency and throughput under a synthetic load test and record the baseline
Your checkpoint
A deployment manifest set with canary routing, rollback policy, and a load-test report showing latency percentiles
MLOps engineering practice for machine learning lifecycle and deployment
Define service-level indicators for prediction quality, latency, and error rate, and encode them as alerts
Implement drift detection on input features and on prediction distributions with a documented threshold
Build a monitoring dashboard that shows traffic, latency, error rate, and drift status on one screen
Simulate a drift event and confirm the alert fires and the runbook steps resolve it
Your checkpoint
A monitoring dashboard plus an alert rule set and a tested runbook for one simulated drift incident
MLOps engineering practice for machine learning lifecycle and deployment
Automate retraining triggered by drift alerts or a schedule, reusing the week-1 pipeline
Add a champion-challenger evaluation that gates promotion on offline and shadow metrics
Implement lineage tracking from dataset version to deployed model version
Audit the pipeline for security basics: secrets management, image scanning, and least-privilege access
Your checkpoint
An automated retraining-and-evaluation workflow with lineage records and a short security audit checklist
MLOps engineering practice for machine learning lifecycle and deployment
Compare batch, real-time, and streaming serving patterns and justify a choice for a stated latency and cost budget
Implement a fallback path that serves a cached or default prediction when the model service is unavailable
Design a cost model for training and inference at a stated request volume and identify the dominant cost driver
Review the whole system against common MLOps failure patterns and list concrete remediations
Your checkpoint
A one-page architecture decision record covering serving pattern, fallback strategy, and cost model
MLOps engineering practice for machine learning lifecycle and deployment
Consolidate the eight weeks of artifacts into a single repository with a clear README and run instructions
Run a timed mock exam focused on lifecycle and deployment scenarios, then review every missed item against the source material
Rewrite the weakest three explanations from memory without consulting notes
Perform a final end-to-end rehearsal: commit, build, test, deploy, monitor, and roll back in one sitting
Your checkpoint
A complete, self-contained portfolio repository plus a scored mock-exam review sheet listing corrected weak areas
Milestones
Week 2: You can rebuild and deploy the model service from a clean checkout using only the README, and CI passes without manual steps
Week 4: You can demonstrate a canary release and an automatic rollback in under ten minutes, with latency numbers to justify the rollout size
Week 6: A simulated drift alert triggers retraining and champion-challenger gating without manual intervention, and lineage links the deployed model to its dataset version
Week 8: The mock exam score is at or above your target pass margin and every missed item has a written correction in the review sheet