Discord has migrated its messages database from Cassandra to ScyllaDB due to scalability and maintenance issues. The Cassandra cluster suffered from hot partitions, leading to unpredictable latency, and required constant maintenance.
- ScyllaDB offered better performance, faster repairs, stronger workload isolation, and no garbage collector.
- Discord also developed data services to control the amount of concurrent traffic to hot partitions, wrote in Rust for its fast concurrency and safety, and implemented consistent hash-based routing.
- The migration of trillions of messages took nine days instead of three months after rewriting the data migrator in Rust.
- ScyllaDB has significantly improved tail latencies and unlocked new product use cases.
