2 min lesson
Cost and the agent-data boundary
Work through the cases in "Cost and the agent-data boundary", pairing each signal with the move that fits.
Step 1 of 2
Cost and the agent-data boundaryThe two clauses generic candidates skip
Right-size clusters and prefer spot for backfills; reserve on-demand for SLA-bound jobs.
Storage layout is the cheapest lever: partition + compact + Z-order before you reach for bigger compute.
Attribute spend per asset/team so cost has an owner, not just a monthly surprise.
Cursor's data feeds agent and model capabilities, so PIIPersonally Identifiable Information. Data that can identify a person (names, emails, SSNs); regulated and sensitive. Press Enter for the full definition. handling is a first-order design input, not a compliance afterthought.
Mask or tokenize PIIPersonally Identifiable Information. Data that can identify a person (names, emails, SSNs); regulated and sensitive. Press Enter for the full definition. before it crosses into training-eligible tables; enforce with Unity Catalog RBACRole-Based Access Control. Granting permissions by role rather than configuring each person individually. Press Enter for the full definition./ABAC.
Retention and audit on sensitive product + agent data; be able to prove who touched what.
When you reach a fork - Kafka vs. direct writes, exactly-once vs. idempotent at-least-once - state the trade and pick, then say what would change your mind. “I'd take at-least-once with idempotent merges because exactly-once doubles the coordination cost for a guarantee an idempotent key already gives me. If duplicates were unrecoverable downstream, I'd reconsider.” That is the autonomy signal the founding-engineer hire is graded on.
Don't design the mature 50-engineer platform. The role is one of Cursor's first data platform engineers, so a v1 that ships and a clear migration path beat a perfect end-state you can't staff. If you draw a five-tool best-of-breed stack with no build-vs-buy reasoning, you've failed the pragmatism filter.
Learn more
Optional practice
Practice: Cost and the agent-data boundary
QFor billions of Cursor events/day landing in bronze, why is at-least-once delivery with idempotent merges usually the right default over exactly-once?