Modern database design for SSDs should shift from single-system durability to distributed approaches. Key insights include: sizing RAM caches for ~30-second hot sets (based on updated Five Minute Rule calculations), optimizing for 32kB transfer sizes to balance throughput and IOPS, committing transactions to distributed logs across availability zones instead of local WALs, and leveraging high-quality hardware clocks for strongly consistent scale-out reads. The relational model, ACID properties, and SQL remain valuable, but durability and high availability should be distributed concerns rather than single-system optimizations.
Sort: