You need a pipeline to copy time-series transaction data into BigQuery for analysis. Dataset initial size 1.5 PB, grows 3 TB/day; thousands of transactions are updated hourly with new status. Data is heavily structured and used for ML. To maximize performance and usability, which two strategies do you adopt?

Denormalizing the data is crucial for BigQuery's performance, especially with large, structured datasets used for ML. BigQuery is a columnar store optimized for analytical queries, and denormalization reduces joins, which are expensive at scale, leading to faster query execution and simplified data models for ML. Appending status updates instead of using UPDATE statements is also key. BigQuery is append-optimized; UPDATE operations are resource-intensive and can lead to performance degradation and increased costs, especially with frequent, small updates on a 1.5 PB dataset. Appending new status records, perhaps with a validto timestamp on the old record, is a common pattern for handling slowly changing dimensions in BigQuery. Preserving the structure (normalization) would hinder performance. Using UPDATE to reduce size is incorrect; it increases processing. External Avro files would be slower than native BigQuery tables for analysis.