Google ML Engineer Course Notes

Start by deciding whether machine learning is actually the right solution. ML fits patterns that can be learned from data, such as prediction, ranking, anomaly detection, clustering, or language generation; deterministic rules, simple reporting, or missing business goals usually need a simpler design first.

• Starting with model selection before defining the prediction target, label, baseline, and business metric.

• If the problem is a known common task like translation, consider a prebuilt API before custom training.

Data engineering for ML is about making training data match production reality. Training-serving skew happens when features are computed one way during training and another way during serving, so the model receives inputs in production that differ from the data it learned from.

• Allowing temporal leakage by training on fields that would not exist at prediction time.

• If the same feature must be reused consistently for training and serving, choose Feature Store or feature views.

Model building starts with the amount of control required. AutoML Tabular is appropriate for structured data when speed and reduced modeling effort matter, while custom training fits specialized architectures, custom loss functions, unusual preprocessing, or framework-specific code.

• Using AutoML when the scenario requires custom architecture, custom loss, custom containers, or framework-specific training code.

• If the task is structured data and speed matters, compare AutoML Tabular before custom training.

MLOps turns model work into repeatable production practice. Pipelines define the steps for data preparation, training, evaluation, registration, and deployment so releases do not depend on manual notebook execution.

• Treating notebooks as the production workflow instead of moving repeatable steps into pipelines.

• If the question asks for repeatable ML workflow orchestration, choose Pipelines.

Deployment choice follows prediction latency and data volume. Online inference serves low-latency requests for applications, while batch prediction scores large datasets asynchronously when immediate response is not required.

• Choosing online inference for large offline scoring jobs that do not need immediate predictions.

• If the data is already in BigQuery and scoring can run offline, choose batch prediction with BigQuery integration.