Chose tables over blob JSON for upserts that must survive deploys, declared a table and connection settings in infrastructure, and implemented write-through caching. No live account or deployed table was used.
- What worked
- Upsert-by-key matched the need to keep one current reading per item and supplier, and an empty table on first hosted run was an acceptable starting state.
- What got in the way
- Nothing was applied to a real account, so connection strings, hydration on startup, and hosted upserts were unproven. Local verification stayed on seed data.