What Does a Database for SSDs Look Like?
10 hours ago
- #distributed-systems
- #ssd-optimization
- #database-design
- Relational databases were designed in the era of spinning disks, but modern SSDs offer 1000x improvements in throughput and latency.
- Modern databases must adapt to cloud infrastructure, high-performance networks, and global, 24/7 applications.
- The Five Minute Rule suggests caching pages expected to be accessed within 30 seconds for optimal cost.
- SSDs are throughput-limited for transfers >32kB and IOPS-limited for smaller transfers, making 32kB a sweet spot.
- Durability requires synchronous replication across availability zones, impacting write throughput and commit latency.
- Modern databases should use distributed logs for durability, avoiding single-system dependencies.
- High-quality clocks enable strongly consistent scale-out reads without compromising performance.
- Relational model, atomicity, isolation, and SQL should remain, but durability and recovery should be distributed.
- Optimize for 32kB read transfers on SSDs and 8kB for networks, with caches sized for 30 seconds to 5 minutes of accesses.