Design a URL shortener has been the same question since 2019.
The question did not expire. Your answer did.
Design Twitter. Design a rate limiter. Design a payment system.
Same questions. New rubric.
For years, this was enough:
- Draw a load balancer
- Add a cache
- Add a database
- Say "we scale horizontally"
- Collect offer
Run that answer today and you get a thank you email.
And no, the fix is not learning all 30 tools in that image.
Naming Kafka, Istio and Cassandra in one breath is not a design.
I read interview writeups every week. I sit with people right after they walk out of these rounds. The pattern is not subtle anymore.
Every candidate now gives the same AI polished answer. Same boxes. Same order. Same words.
So interviewers changed what they listen for. Here is how the round actually runs right now:
1. They interrupt you.
- You say: we will shard the database.
- They ask: shard on which key?
Then: one tenant is 40 percent of your traffic. Now what?
Then: how do you rebalance without downtime?
- Saying "we will shard" is free. A memorised answer survives one follow-up. Not three.
2. Cost is on the rubric.
- Multi region Cassandra for a 1000 user app is a reject.
- An LLM call on every request with no caching is a reject.
Not because they fail. Because you never asked what they cost per month.
3. AI moved into the normal loop.
- Design a recommendation system is not collaborative filtering anymore.
- It is embeddings, a vector store, retrieval, and a bill that does not kill the company.
You do not need to be an ML engineer. You need one answer ready: what happens when the model provider times out at peak traffic.
4. Failure modes are the round.
- The diagram is the setup. Not the answer.
- Redis dies at 12:00:01. What breaks first?
- Your Kafka consumer lags four hours. Drop or replay?
- Primary fails over mid transaction. Who owns that money?
Two components explained deep beat five drawn shallow. Every time.
5. The room came back.
- Google and Cisco brought in-person rounds back into the loop.
- Not for culture fit. Candidates were reading answers off a second screen.
- Other companies went the opposite way. They let you use AI in the round, then grade whether you can verify what it handed you.
Ask your recruiter which one you are walking into. Do not guess.
6. In India, the LLD round is still the real filter.
- Machine coding. One sitting. Working code.
- Classes, interfaces, a running demo.
A beautiful diagram scores zero if nothing runs.
The round never rewarded the drawing. It rewarded the person who has been paged at 2am.
System design interviews did not get harder. They got harder to fake.
1September 1, 2026 359
