The choice between LSM tree and B-Tree is one of the most important storage engine decisions. They represent fundamentally different tradeoffs:

LSM Tree (RocksDB, Cassandra, LevelDB):

• Writes are sequential (append to MemTable, flush to SSTable) -> very high write throughput
• Reads may need to check multiple levels -> higher read latency
• Space amplification > 1.0 (old versions exist until compaction)
• Write amplification is high (data rewritten during compaction)

B-Tree (PostgreSQL, MySQL InnoDB, SQLite):

• Writes are random I/O (in-place updates) -> lower write throughput
• Reads traverse a balanced tree (O(log N) disk reads) -> lower read latency
• Space amplification ~1.0 (in-place updates, no dead versions)
• Write amplification is lower

Benchmark both engines and measure: write ops/sec, read p50 latency, and space amplification.

Request:  {"type": "lsm_benchmark", "msg_id": 1, "entries": 100000}
Response: {"type": "lsm_benchmark_ok", "in_reply_to": 1, "lsm": {"write_ops_sec": 200000, "read_p50_us": 50, "read_without_bloom_p50_us": 500, "space_amplification": 1.3}, "btree": {"write_ops_sec": 50000, "read_p50_us": 20, "space_amplification": 1.0}}