1 min lesson
Three consumer classes, three different SLAs
Give a practical answer to this: "Cursor's data platform feeds three distinct consumer classes. Roughly how many events per second should you expect to reason about per 'billion events per day' and why does the consumer split matter for design?"
Step 1 of 2
Three consumer classes, three different SLAsthe part most candidates miss
Cursor's data does not feed one audience. It feeds product analytics, business decisions and the agent/model capabilities themselves. Each has a different tolerance for latency, a different correctness bar and a different security posture.
Feature usage, funnels, experiment readouts.
Tolerates hours of latency; wants breadth and history.
SLA: freshness in hours, high schema stability.
Revenue, growth, board-level metrics.
Correctness is non-negotiable; numbers get quoted externally.
SLA: accuracy and lineage over raw speed.
Training and evaluation data feeding product capabilities.
PIIPersonally Identifiable Information. Data that can identify a person (names, emails, SSNs); regulated and sensitive. Press Enter for the full definition. and licensing constraints are first-order.
SLA: governed access, reproducibility, strict audit.
When you sketch a serving layer, name the consumer class you are designing for and its SLA before you draw a box. "For business metrics I'd optimize for correctness and lineage; for agent training data I'd optimize for governed, reproducible access." Differentiating SLAs unprompted is the fastest way to sound like you have run a real platform.