1 min lesson
The over-scoping trap
Show why this lesson detail matters: "Treat the data platform as a product with internal customers, out loud."
Step 1 of 2
“I scoped to one bronze ingest path and a single silver merge so I could demo a clean, queryable table end to end. I partitioned by event date because that's the dominant filter and I made the merge idempotent so a retry is safe. What I'd harden next is compaction for small files and a freshness alert - I left the merge behind an asset boundary so dropping it into Dagster is a small step.”
Treat the data platform as a product with internal customers, out loud. Ask “who queries this and how often?” early and let the answer drive your partitioning and serving layout. That framing is exactly the product sense the onsite is built to detect and most candidates skip it.
The most common onsite failure is starting an ambitious architecture and demoing nothing. A working slice with two honest trade-offs beats an elaborate design with no running code.