The platform bill doubled. Where do you look?
Whether the candidate reasons about drivers rather than about totals, and whether they attribute before optimising.
The situation behind the question
Interviewers ask this because it happened to them.
Warehouse spend has grown sharply over a quarter with no change in headcount and no obvious new workload. Every individual query the team looks at seems reasonable.
A strong answer
Flags
Green flags
- Thinks about cost as a set of drivers with different remedies, not as a single number to reduce.
- Attributes before optimising, and treats the first attribution report as the intervention.
- Connects scan cost back to physical layout and column selection, which are design decisions made earlier.
- Notices refresh cadence as a driver that nobody chose deliberately.
Red flags
- Reaches for a cheaper storage tier when the driver is scan volume, and does not account for retrieval cost from the colder tier.
- Proposes reducing cluster size without knowing which workload is responsible.
- Cannot name more than one cost driver.
- Treats cost as an infrastructure problem rather than as a consequence of layout, modelling and schedule.
Follow-ups
Where the conversation goes if the first answer holds up.
- How would you attribute shared warehouse compute to teams, and what does your method get wrong?
- A dashboard refreshes every five minutes and is read once a day. Whose decision is that, and how would you find it?
- Which cost driver is the hardest to reverse once a hundred queries depend on it?