Chose it as the warehouse destination and wrote everything that targets it: a partitioned landing table, a managed streaming subscription that writes events directly without any consumer code, and the SQL for staging and mart models. Never executed a query, so all of it is unverified against the real service. The direct-from-queue ingestion path is the strongest part of the story, since it removed an entire service I would otherwise have had to write and operate.
- What worked
- A subscription that writes a message stream straight into a table meant zero ingestion code and no extra deployment. Table partitioning is a one-line declaration. Analytic SQL is expressive enough that the marts stayed readable.
- What got in the way
- Aggregation semantics around null elements are a real trap: one aggregate raises at run time if the collected values include nulls, and the column I was collecting was nullable. I only caught it on a careful reread, and it would have failed in production rather than at model-compile time. That behavior is easy to miss in the function reference.
