NotesSystem
Cassandra
Also: Apache Cassandra
A wide-column database descended from Dynamo's design: leaderless replication with per-query tunable consistency and last-write-wins per cell.
Design choices
- Leaderless Replication: any node coordinates a request and fans out to the key's replicas.
- Quorum: consistency levels per query:
ONE,QUORUM,LOCAL_QUORUM(majority in the local datacentre),EACH_QUORUM,ALL. - Last Write Wins: conflicts resolve per cell by write timestamp.
- Hinted handoff: coordinators keep hints for a down replica up to
max_hint_window(3 h default); best effort. The consistency level counts acknowledgements from the key's own replicas, and onlyANYaccepts a stored hint in place of one, so Cassandra's quorums are strict. - Read Repair: blocking, foreground repair on mismatched reads; background probabilistic repair was removed in 4.0.
- Anti-Entropy:
nodetool repairwith Merkle trees; incremental by default.
Notable trade-offs
- Repair is the operator's job: run it at least every 7 days with the default 10-day
gc_grace_seconds, or deleted data can resurrect. - LWW means concurrent updates to the same cell lose data; model data so cells are written once or contention is rare.