A retail client needs BI dashboards for 400 concurrent users and a data science platform over the same data. One lakehouse or a lakehouse plus a serving layer?

What the interviewer is testing

Whether you can assess the "one system for every workload" claim against the specific workload characteristics.

The workloads are genuinely different

BI at 400 concurrent users needs sub-second response on aggregate queries, high concurrency, predictable latency, and small result sets. Data science needs full-table scans, arbitrary compute, Python and notebook access, and tolerates minutes.

The lakehouse foundation serves the second superbly: open formats on object storage, any engine, cheap storage, no copies.

For the first, high-concurrency low-latency BI directly on object storage remains the weakest point. Query planning against many files, cold metadata and the absence of specialised caching mean p95 latency is usually seconds rather than sub-second, and 400 concurrent users amplify it.

The recommendation

Lakehouse as the foundation, with a serving layer for BI. Concretely:

• Open table formats on object storage as the single source of truth
• Data science, ML and ad hoc analysis read directly from it
• Curated aggregates and dimensional marts materialised into a serving store tuned for concurrent aggregate queries
• One transformation pipeline producing both, so definitions do not diverge

The serving store holds derived data only, which keeps the copy governable — it is regenerable and never authoritative.

The alternative worth acknowledging

If the platform's SQL engine has a caching and result-serving tier and the dashboards are predominantly against pre-aggregated marts, a single system may well be sufficient. That should be established by benchmarking at the real concurrency with the real query mix, not by vendor claim.

What a strong answer adds

Naming the test that decides it: load-test the dashboard query mix at 400 concurrent users against the candidate architecture with production-scale data. This is a one-week spike and it settles an argument that otherwise runs for months.

And noting that a semantic layer above both keeps metric definitions consistent regardless of which store answers.

Common weak answers

Choosing a single lakehouse on principle. Recommending a traditional warehouse and re-creating the two-copy problem.