System Design Interview Preparation
System design interviews aren't testing whether you know the name of every database and queue on the market — they're testing whether you can navigate ambiguity, make explicit tradeoffs, and justify them under scrutiny. Here's a preparation approach that holds up under real interview pressure.
Start by scoping the problem, out loud
The most common failure mode is jumping straight to a diagram. Before drawing anything, state your assumptions about scale (requests per second, data volume, read/write ratio), clarify which requirements are hard (e.g. strong consistency for payments) versus soft (e.g. eventual consistency is fine for a feed), and explicitly list what's out of scope. This alone signals seniority before you've designed anything.
Design in layers, not all at once
Work outward from the API contract, to the core data model, to how services communicate, to how you'd scale each piece under load. For each layer, name the specific decision and the alternative you rejected: "I chose Kafka here for durable event replay and consumer scaling; I rejected direct synchronous calls because a downstream outage shouldn't block the write path." That contrast is what interviewers are listening for.
Have real numbers ready for common patterns
Be able to reason about caching strategy and invalidation (cache-aside with a TTL is the default most candidates should be fluent in), database indexing choices for the access patterns you just described, and how you'd shard a dataset that's outgrown a single instance. You don't need to memorize exact throughput numbers, but you should be able to reason about orders of magnitude and where a bottleneck will appear first.
Handle the failure-mode follow-up
Almost every system design interview ends with "what happens when X fails?" — a downstream service times out, a message is delivered twice, a region goes down. Prepare concrete answers involving retries with backoff, idempotency keys, circuit breakers, and graceful degradation, tied to the specific design you just presented, not generic buzzwords.