ScyllaDB

An optimised version of Cassandra, written in C++ to get rid of java gabage collection

Use LSM Tree (Log Structured Merge Tree) which remove the need of lock. DynamoDB would use B+ tree and Lock requires.

Support lightweight ACID via Paxos/Raft. ScyllaDB has "shard-per-core" (1 shard = 1 CPU core) architecture that can support high throughput of write in parallels.

Use case

  • Very very high write/read throughput needed
  • lightwieght ACID is acceptable

Cons

  • Consistency model is not as mature as DynamoDB or SQL
  • Overhead to setup
  • Require some network to communicate — use postgres if this is a requirement