Consistency, Availability, and Latency Tradeoffs

Strong Consistency in Relational Systems
Relational databases with a single primary (the server handling all writes) provide strong consistency: every read sees the most recent write. The trade-off is availability during failover, typically 10-30s. Within a region, commits take 1-5ms. Cross-region uses two-phase commit (2PC): coordinator asks all to prepare, commits if all agree, adding 50-150ms.
High Availability in NoSQL Systems
NoSQL uses leaderless replication (any node accepts writes) with quorum operations. Quorum means majority: with 3 replicas, 2 must acknowledge. Writes continue if one node fails. The cost is eventual consistency (clients may temporarily see different values). Latencies stay 1-5ms.
Strong Consistency
1-5ms writes, 10-30s failover
vs
High Availability
1-5ms writes, no downtime
Distributed SQL: Bridging the Gap
Distributed SQL provides ACID with horizontal scaling via consensus (nodes vote to agree). Within region: 10-20ms. Cross-region: 50-150ms. Some offer tunable consistency: strong for critical ops, eventual for less critical.

💡 Key Takeaways

✓Single-primary relational provides strong consistency with 1-5ms local writes but 10-30 second unavailability during failover

✓Leaderless NoSQL with quorum (majority agreement) achieves high availability with single-digit ms latency by accepting eventual consistency

✓Two-phase commit (2PC) coordinates distributed transactions but adds 50-150ms latency and fails if the coordinator crashes

✓Distributed SQL bridges the gap: ACID guarantees with horizontal scale, but 10-20ms regional and 50-150ms cross-region write latency

📌 Interview Tips

1When discussing consistency trade-offs, quantify the latency impact: strong consistency adds 10-20ms per region hop via consensus protocols

2Explain quorum clearly: with 3 replicas requiring 2 for writes and 2 for reads, you guarantee reading the latest write (2+2 > 3)

3For high-availability requirements, mention that leaderless systems avoid single points of failure but push conflict handling to the application

← Back to Relational vs NoSQL Overview