System design interviews in 2026: what actually gets probed

System design rounds have quietly changed. With AI-assisted coding now assumed, loops shifted weight onto judgment: trade-offs, failure modes, and cost. We tagged 900 follow-up questions from mock sessions to see where interviewers actually push.

The top three probe areas

1) Failure and degradation (31% of follow-ups): 'What happens when this queue backs up?' 2) Data consistency (24%): 'Walk me through a write conflict.' 3) Cost and capacity (17%): 'What does this cost at 10x traffic?' Classic topics like load balancing barely register anymore — they're table stakes.

Where candidates hand-wave

The single most common weak moment: quoting a technology as the answer ('we'd use Kafka') without stating the property it buys you (ordering per partition, replay, backpressure). Interviewers almost always follow up, and the follow-up is where scores are decided.

How to practice this

Run a system-design mock and let the AI interviewer probe. Then read the transcript and highlight every sentence where you named a tool. For each one, write the property you were actually buying. Next session, say the property first and the tool second. This single inversion moved our users' technical scores more than any other drill.