LLM infrastructure questions have entered the general loop
'Design a system that serves an LLM' used to be an ML-role question. In 2026 it shows up in ordinary backend loops. Here's the vocabulary you're expected to have.
Cost per request is the new scalability question
Senior system design loops in 2026 grade cost reasoning explicitly. 'It scales' is no longer a complete answer if it scales at $40k a month.
The ops questions most candidates skip: observability, rollbacks, on-call
Deployment strategy and incident thinking are no longer bonus topics in system design rounds. Senior candidates who skip them are leaving signal on the table.
Designing retrieval: how to handle the RAG question
'Make the chatbot answer from our documents' is 2026's most common AI system design prompt. The pipeline is simple; the grading is in the failure modes.
The classics still decide loops: queues, caches, and consistency
AI questions get the headlines, but most system design rejections in 2026 still come from fumbled fundamentals. The core topics and the depth actually required.
Same question, three answers: system design at mid, senior, and staff
Interviewers use one prompt and calibrate level from the shape of the answer. What actually distinguishes L4, L5, and L6 responses to identical questions.