Compaction is the background process that keeps the LSM tree healthy. Without it, reads degrade because more and more SSTables accumulate, each requiring a disk seek.

The compaction algorithm (leveled compaction):

1. Trigger: when L0 exceeds 4 SSTables, or when a level exceeds its size limit
2. Select: choose SSTables from the source level whose key ranges overlap with the target level
3. Merge: read all selected SSTables, merge-sort by key, deduplicate (keep newest version of each key)
4. Write: write new sorted, non-overlapping SSTables to the target level
5. Cleanup: delete the old source SSTables

Key metric - write amplification: a single user write may be rewritten 10-30x across levels as data is compacted. This is the fundamental cost of LSM trees. RocksDB's leveled compaction has write amplification of approximately 10 * (number_of_levels - 1).

Request:  {"type": "lsm_compact", "msg_id": 1, "source_level": 0, "target_level": 1}
Response: {"type": "lsm_compact_ok", "in_reply_to": 1, "sstables_merged": 4, "new_sstables": 2, "keys_deduplicated": 500, "bytes_read": 16000000, "bytes_written": 14000000, "write_amplification": 2.3}

Request:  {"type": "lsm_compact_status", "msg_id": 2}
Response: {"type": "lsm_compact_status_ok", "in_reply_to": 2, "pending_compactions": 0, "total_compactions": 5, "total_bytes_compacted": 100000000}