Designing data preparation and processing systems
This domain covers ingestion, transformation, feature engineering and data quality for both classical ML and retrieval-augmented GenAI. The exam tests whether you pick the right processing service and avoid leakage and training-serving skew.
Core rules
- Match the service to the data shape. BigQuery for set-based ELT at scale; Dataflow (Apache Beam) when one codebase must serve streaming and batch; Dataproc when an existing Spark or Hadoop job has to move; Datastream for change data capture from operational databases; Cloud Storage as the immutable landing zone.
- Land raw, curate separately. Keep raw objects untouched, write transformed tables with versioned schemas, and retain the transformation code.
- Split before you transform. Fit normalization, encoders and imputation on training data only; split time series by time and grouped records by entity.
- Prevent skew with one feature definition. Generate features through Vertex AI Feature Store or a shared transformation module rather than reimplementing them in the serving path.
- Prepare GenAI data deliberately. Chunk documents with overlap, embed with a Vertex AI embedding model, deduplicate near-identical chunks, and keep metadata for filtering and citations in Vector Search.
- Handle PII before training using the Sensitive Data Protection (DLP) API and separate, access-controlled buckets.
Worked example
Clickstream events must feed a fraud feature ("events in the last 5 minutes"). Reasoning: events arrive continuously, so publish to Pub/Sub, aggregate in Dataflow over a five-minute window, write to BigQuery for training and to Feature Store for online lookups. The same Beam aggregation serves both paths, so the feature cannot drift between training and serving.
Traps
- Random splits on time-ordered data, leaking the future into training.
- Computing statistics over the full dataset before splitting.
- Standing up a Dataflow job for a small nightly file that BigQuery or Dataform handles.
- Writing TFRecords without shuffling, producing ordered, biased batches.
- Chunking RAG sources so coarsely that retrieval cannot isolate the answer, or so tightly that sentences lose context.
- Skipping deduplication, which inflates evaluation scores and over-weights repeated content.