MLOps & Model Deployment Basics
A large share of models built in Malaysian organisations never leave the notebook they were trained in, and a good number of the ones that do are quietly wrong within a year because nobody watched them. The gap is operational rather than mathematical. This day covers the path from a trained model to a service someone depends on, and the monitoring that tells you when the world has moved away from what the model learned.
Programme Agenda
Why Models Stall Before Production
The handover problem between data science and engineering, environment differences, the missing owner, and the feature that was computed in a notebook and exists nowhere else.
Reproducibility
Pinning data, code, parameters and environment so a result can be reproduced six months later. Experiment tracking, and the discipline of recording what was tried and rejected.
Packaging and the Model Registry
Serialising a model with its preprocessing, versioning it, and promoting through stages. Why the preprocessing must travel with the model rather than live in the calling application.
Serving Patterns
Batch scoring, real-time endpoints and embedded models, compared on latency, cost and operational load. Choosing based on how the prediction is consumed.
Deployment and Rollback
Shadow deployment, canary release and champion-challenger comparison. Keeping the previous model available and the conditions that should trigger a rollback.
Monitoring: Data and Concept Drift
Watching input distributions shift and the relationship between inputs and outcome change. Setting up drift detection when ground truth arrives weeks after the prediction.
Retraining
Scheduled against triggered retraining, the data window to use, validating that the new model is genuinely better, and automating the parts that are safe to automate.
Ownership and Governance
Who owns a model in production, the documentation that should exist, audit expectations where a model affects customers, and decommissioning a model nobody uses any more.
Learning Outcomes:
Identify the organisational reasons models stall before production
Make a training run reproducible in code, data, parameters and environment
Package a model with its preprocessing and register it with a version
Choose a serving pattern matched to how predictions are consumed
Deploy with a rollback path and defined rollback conditions
Monitor for data and concept drift including delayed ground truth
Decide between scheduled and triggered retraining and validate the result
Assign ownership and documentation for a model in production
Duration: 1 Day (8 Hours)
Training Hours: 9:00 AM to 5:00 PM
Level: Beginner
Training Mode: Physical, Online, or Hybrid
HRD Corp SBL-KHAS Claimable
Certificate of Completion included
Frequently Asked Questions
More in Data Engineering and Machine Learning
- Cloud Platform Foundations (AWS and Azure) · Beginner, 1 day
- Data Engineering and Pipelines · All levels, 2 days
- Data Science and Machine Learning Fundamentals · Beginner, 1 day