1.1.2. The Experiment-to-Production Gap

💡 First Principle: A model that works in a Jupyter notebook is to production ML what a sketch on a napkin is to a skyscraper. The gap isn't about model quality—it's about everything the notebook doesn't address: reproducibility, scalability, monitoring, security, and cost.

Most ML projects fail not because the model is bad, but because the team can't operationalize it. A data scientist builds a promising prototype in a notebook, training on a sample dataset on their laptop. But moving to production requires answering questions the notebook never asked: How does this model get fresh data every day? What happens when the input data format changes? How do you roll back if the new model version is worse? Who has access to the training data, and is it encrypted?

Consider the journey from experiment to production as crossing five bridges:

Bridge 1: Data Pipeline — Moving from "I downloaded a CSV" to automated, validated, versioned data ingestion (Domain 1).

Bridge 2: Training Infrastructure — Moving from "I trained on my laptop" to scalable, reproducible training on managed compute (Domain 2).

Bridge 3: Deployment Infrastructure — Moving from "I called model.predict()" to serving thousands of predictions per second behind an endpoint (Domain 3).

Bridge 4: Automation — Moving from "I retrained manually" to CI/CD pipelines that automatically test, validate, and deploy model updates (Domain 3).

Bridge 5: Operations — Moving from "I checked accuracy once" to continuous monitoring, alerting, and automated retraining triggers (Domain 4).

The MLA-C01 exam is structured around these five bridges. Each domain tests your ability to cross one or more of them using AWS services. Understanding this structure helps you anticipate what the exam is actually testing behind any given scenario question.

⚠️ Exam Trap: When a question presents a data scientist's notebook workflow and asks how to "productionize" it, the answer almost never involves changing the model. Instead, look for answers that address infrastructure, automation, or monitoring. The exam penalizes candidates who focus on the model when the question is about the system around the model.

Reflection Question: A team has a model achieving 92% accuracy in their SageMaker notebook. They want to deploy it to production. What are the first three non-model concerns they should address?