FDE System Design Interview Scenarios (with What Good Looks Like)
Updated · Tech checked
FDE design rounds use realistic, constraint-heavy scenarios - flaky upstreams, dual systems, PII boundaries, cost caps - not whiteboard URL shorteners. Four worked scenarios with evaluation criteria.
How FDE design rounds differ
Product-company design rounds reward scale gymnastics ("design Twitter"). FDE design rounds reward constraint navigation: messy data, systems you don't own, security walls, cost ceilings. The rubric usually checks: requirement clarification, failure handling, data integrity, security boundaries, and honest phasing.
Scenario 1 - The flaky ERP integration
"Customer's ERP exposes a REST API that times out ~10% under load and occasionally delivers duplicate webhooks. Build order sync." Strong answers: idempotency keys on consumer; exponential backoff with jitter + budget; dead-letter queue with reconciliation report; watermark-based incremental pulls as webhook fallback; explicit answer on ordering (don't assume it). Red flag: trusting webhooks alone.
Scenario 2 - The permission-aware assistant
"Support assistant that answers from docs. Different agents see different document sets. Design it." Strong answers: retrieval filtered by ACL pre-retrieval, not post-filtering of results; citation of source doc IDs; audit log of queries→docs served; evaluation plan including unauthorized-document probes. Red flag: post-hoc filtering ("we'll just hide results") - leaks happen between retrieval and filter.
Scenario 3 - The dual-run cutover
"New pricing engine must replace a legacy rules service without downtime. Design the rollout." Strong answers: shadow mode (new engine computes, old engine decides) with diff dashboards; expand/contract data migrations; flag-based percentage cutover with instant rollback; reconciliation job and sign-off criteria before decommission. Red flag: big-bang with "we tested."
Scenario 4 - The cost-capped AI workflow
"Document summarization for 50k docs/month; latency < 3s p95; monthly AI budget fixed." Strong answers: tiering (cheap model first, escalate on low confidence); caching by content hash; batch windows for non-interactive docs; token budgets + per-workflow cost metric; degradation plan (queue, don't fail). Red flag: premium model everywhere, cost discovered at invoice.
How to practice
Run each scenario aloud in 25 minutes: 5 clarifying, 10 designing, 5 failure/security pass, 5 phasing. Then compare against the rubric points above. Deeper prep in the interview guide.
Continue: Customer-facing questions · Take-home prep