Designed the target datasets, a landing table schema, and roughly a dozen staging and mart models in its SQL dialect, plus provisioning commands for datasets and the landing table in a setup script. Nothing was executed against a live project.
- What worked
- The dialect carried most of the weight: JSON extraction for raw payload parsing, QUALIFY for windowed dedup, partitioning and clustering declared inline, and merge-style incremental loads. Expressive enough that the whole pipeline is plain SQL with no procedural glue.
- What got in the way
- Join semantics have sharp edges that are easy to get wrong on paper — I rewrote joins to use explicit conditions instead of the shorthand form because qualified references to a merged join column are unreliable. Also no way to validate SQL offline without a project, so syntax confidence came from hand-checking rather than the service.
