Performance & Fault Tolerance: Building a Modern Database
summary
ScyllaDB claims x10 throughput over Cassandra by being written in C++ and running shard-per-core so each core owns CPU IO and storage, avoiding GC pauses that cause 10s stalls in Java systems. A large user migrates from Cassandra to ScyllaDB and reduces 177→72 nodes and cuts P99 from up to 125ms down to about 15ms, supporting millions of ops/sec with sub-millisecond P99 under typical workloads. Elastic scaling uses Raft-based tablets to add nodes in minutes with 80–90% storage utilization, and it runs as DBaaS, on cloud, or on prem.
AI-written summary · about
chapters (12)
- 0:00Introduction and welcome
- 1:25Discord’s migration journey
- 8:31New generation databases
- 12:36Modern hardware and CLAB
- 15:26What is CLAB
- 20:01Deep dive into architecture
- 25:20Consistency levels explained
- 28:09Multiple data centers
- 33:19Quick wins lab demo
- 35:19Elasticity in CLAB
- 38:13When to use CLAB
- 39:20Q&A and closing
related talks
44mScyllaDB where Cassandra meets Redis!
35mHow to Save on Scaling by Moving from Cassandra to Scylla DB
30mOvercoming Distributed Databases Scaling Challenges with Tablets
40mLatency, Throughput & Fault Tolerance: Building a Modern Database
26mA deep dive and comparison of Python drivers for Cassandra and Scylla
01h 02mBreak Out of Redis
25mCache? Where we're going, we don't need cache!
11mScyllaDB Tablets in Action