The Backend Engineering Show with Hussein Nasser Why Discord Moved from MongoDB to Apache Cassandra, Let us Discuss
10 snips
Oct 9, 2020 AI Snips
Chapters
Transcript
Episode notes
Index Size Determines DB Performance
- Discord's MongoDB setup hit unpredictable latency when indexes no longer fit in RAM.
- When indexes exceed memory, databases shift from CPU-bound to I/O-bound and performance degrades dramatically.
Partition Early To Avoid Memory Thrashing
- Partition or shard indexes to keep index working sets small and in memory.
- Avoid single huge indexes; design partitions by access pattern like channel-id to limit memory growth.
Cassandra's Partitioned Data Model Fits Chat
- Cassandra offers built-in partitioning and a composite key model (partition, primary, clustering).
- That model routes queries to the node holding the partition, reducing cross-node scanning for channel queries.