Compaction
Merging LSM SSTables to bound read amplification.
Origin
Comes from the LSM-tree paper (O'Neil et al., 1996). RocksDB documents the modern tiered vs levelled trade-offs more clearly than any other reference.
Where it shows up in production
- RocksDB Levelled compaction default; SSTables in level N are merged into N+1 in the background.
- Cassandra Three strategies: STCS (size-tiered), LCS (levelled), TWCS (time-window) — pick by workload.
- ScyllaDB Improves on Cassandra's compaction by parallelising across shards.
On Semicolony
Sources & further reading
Found this useful?