What will users see
Once data is replicated, interviewers probe consistency. The question is what a user observes after a write, especially when reads can hit a stale replica. A good answer matches the consistency level to what each feature actually needs.
A spectrum not a switch
- Strong consistency means a read always reflects the latest write.
- Eventual consistency means replicas converge after some delay.
- Read your writes means a user at least sees their own changes.
- Monotonic reads mean a user never sees time go backward.
Match level to feature
- A bank balance wants strong consistency.
- A like count tolerates eventual consistency.
- A user profile edit wants read your writes.
The cost
Stronger consistency usually costs latency and availability, since it may require coordination across replicas. State which features pay that cost and which relax it, rather than applying one level everywhere.
Key idea
Consistency is a spectrum, so match each feature to the weakest level it can tolerate and accept the latency and availability cost only where strong guarantees are truly needed.