Design a cloud database write path and recovery

System Design (Engine-level): Write Path + Crash Recovery

Design a core subsystem for a cloud-native relational database (Aurora-like) where compute is separated from durable distributed storage.

Goal

Support transactional writes with:

• high throughput
• low commit latency
• crash recovery
• strong durability guarantees (clearly specify what guarantees)

Requirements / prompts

1. Write path : Describe how an UPDATE/INSERT flows from compute to durable storage. Where do you place the log (WAL)?
2. Commit protocol : When does a transaction commit succeed? What acknowledgements are required?
3. Replication & consistency : How many replicas, what quorum rules, and how do you handle network partitions?
4. Crash recovery : If the compute node crashes, how does a new node recover state and resume service? What data structures/checkpoints exist?
5. Write amplification : Identify sources (WAL, page rewrites, compaction) and propose reductions.
6. Scalability : How do you scale storage and compute independently? Discuss sharding, rebalancing, and hotspot handling.
7. Observability : What metrics and logs would you add to detect replication lag, redo backlog, and tail latency?