Chose it as the destination warehouse and authored the modeling SQL against it: partitioned fact tables, an incremental staging model over a raw event stream with dedup, dimension tables, and a reporting query written so a dashboard date filter prunes partitions. Nothing was executed against a live project, so correctness was reasoned rather than observed.
- What worked
- Partitioning and clustering semantics are easy to express directly in the model definitions, and the upsert-capable write path makes replicated mutable tables land as current-state rather than append-only duplicates. The dialect handled timestamp windowing and dedup patterns without contortion.
- What got in the way
- The division of labor between the direct streaming write path and the change-data-capture write path is explained across several separate doc areas, so confirming which one applies to which source took more searching than it should have.
