API design, data modelling, concurrency and distributed systems - practise the rounds that decide a backend offer.
Backend loops are the most standardised in the industry, which cuts both ways: you know exactly what is coming, and so does every other candidate. The differentiator is depth. Everyone can say "add a cache". Fewer can say which cache, what the invalidation strategy is, what happens on a cache stampede, and what the failure mode looks like when the cache is down.
Expect API design to come up even when it is not a named round. Interviewers use it as a cheap probe: given a feature, define the endpoints, the request and response shapes, the idempotency behaviour, the pagination contract and the error semantics. Vague answers here undercut an otherwise good design.
The concurrency round is where mid-level candidates most often lose points. Knowing that a race condition exists is table stakes; being able to name the exact interleaving that causes it, and choosing between a lock, an atomic operation, an optimistic version check or a queue with a reason, is the signal.
45 min
Medium to hard problem, with complexity discussion.
45 min
Endpoints, schema, indexes, idempotency, pagination and error contracts.
60 min
Scale a service: caching, sharding, queues, consistency, failure handling.
60 min
Class structure, SOLID, extensibility. Heavily weighted at Indian product companies.
45 min
Ownership, on-call incidents, and technical decisions you defended.
Numbers and failure modes. An average answer lists components. A strong one estimates traffic and storage, picks the bottleneck those numbers imply, explains what happens when each component fails, and names the trade-off it accepted - usually consistency, cost or latency.
Enough to choose a store and defend it. You should be able to explain why you picked relational over document, what your primary access patterns are, which indexes support them, what the write amplification cost is, and how you would shard when one node stops being enough.
Less often as a dedicated round. US big tech folds object-oriented design into the coding or HLD round. Indian product companies frequently run a full machine coding or LLD round of 60 to 90 minutes, and it carries the same weight as system design.
Run a system design mock and find the bottleneck before an interviewer does.