Making Cassandra Faster in Cloud-native architecture
summary
Salesforce ports Cassandra 3.11 into Kubernetes with strict FIPS/non‑FIPS encryption and a service mesh, measures with nodetool/JMX plus perf, perf record and flame charts, and targets p95 <2 ms while matching data‑center throughput. Findings: native mTLS p99 ≈ 9 s at 100k rps so move to Envoy/Istio; split commitlog to a dedicated SSD volume and enable trickle_fsync to avoid write stalls; enable LC4 compression to cut inter‑node bandwidth 2–3x. Cassandra 4 shows tighter, lower read latencies.
AI-written summary · about
chapters (12)
- 0:00Introduction and use case
- 1:55Deployment and goals
- 3:41Bottleneck identification
- 5:50Performance framework design
- 8:07Encryption overhead analysis
- 10:20Metrics and monitoring
- 15:00Profiling and tooling
- 17:28Analyzing hotspots
- 19:08Optimizing IO and latency
- 27:37Network and compression
- 30:32Cassandra 4 performance
- 33:15Q&A and conclusion
related talks
40mLatency, Throughput & Fault Tolerance: Building a Modern Database
45m8ms/99th write percentile latency - is it fast? (...)
35mReplacing Cassandra: A Digital Transformation for the World’s Largest Digital Identity Network
33mCassandra Performance Tuning Like You've Been Doing It for Ten Years
26mHow Cassandra 4.X Help in Cutting Cost Mission
48m8ms/99th write percentile latency - is it fast?
29mCassandra - an introduction
45mMigrating Mission Critical Workloads from Mainframe to Open...