Zheheng demonstrated strong technical depth and structured thinking during the system design session. He opened with a clear delivery framework to guide the discussion, which was an excellent start.

Areas of strength included:

• Detailed listing of functional and non-functional requirements (with numeric metrics).
• Strong entity modeling with edge cases such as threads.
• Effective use of time—dedicated ~8 minutes to discussing requirements and entities before moving forward.
• Clear RESTful API design for foundational operations like creating channels and user management.
• Thoughtful sequencing of system events: beginning with channel setup before messaging.
• Smart DB decisions: choosing Cassandra for heavy writes while leaving space for SQL alternatives based on partitioning.
• High-level design was structurally sound, with Redis pub/sub to scale channels and handle real-time presence tracking.
• Deep thinking on hot-partition problems (e.g., celebrity chats) using Redis TTL leasing strategy.
• Creative inbox design optimization using a chat_participant_state table to support scalable read/write.
• Consideration of data retention and ordering strategies (Kafka offsets, Snowflake ID generation).
• Introduced session-related device_user_mapping and thought through multi-device notification strategies.
• Covered message deletion workflows for both online and offline user scenarios.

Areas for improvement:

• Grouping functional requirements into broader themes could help improve time management.
• Real-time response for online users is typically defined as < 200ms, not < 500ms—refine your latency targets.
• Improve estimation math (e.g., 1 day ≠ 10⁶ sec; it's ~10⁵).
• Avoid overly generic scalability estimation; focus on peak throughput and bottlenecks instead.
• You could have skipped repetitive API explanations and instead voiced them at a high level to save time.
• Missed briefly touching on message deletion logic and multi-device support in the early stages.

Overall, Zheheng shows solid readiness for senior-level system design interviews, and with some fine-tuning in framing and estimation strategies, should continue to perform strongly across similar questions.

‍