Storage Engine Decision Matrix

Systems Design Framework

Every storage engine embodies trade-offs dictated by the RUM Conjecture: an engine can optimize for at most two of Read amplification, Update (write) amplification, or Memory/space amplification.

🧭Storage Engine Architecture Decision Matrix

Interactive Tool

Storage engines make fundamental trade-offs governed by the RUM Conjecture (Read, Update, Memory/Space). Select your workload profile to explore the optimal storage structure:

Log-Structured Merge Tree (LSM)

Append-Only Multi-Level Sorted String Tables (SSTables)

Optimal Fit

Writes are written sequentially to a Write-Ahead Log (WAL) and memory table (MemTable), then flushed to disk as immutable SSTables. Compaction merges runs in the background. Exceptional write throughput.

Dimension	Characteristics & Trade-off	Impact
Read Amplification	Medium (Bloom-filtered)	Read Path
Write Amplification	Low (Sequential Append)	NAND Wear / Throughput
Space Amplification	Low (Compacted)	Disk Footprint

Notable Production Systems:

RocksDBApache CassandraLevelDBCockroachDB (Pebble)ScyllaDB

---

Architectural Comparison Table

Architecture	Primary Data Structure	Best Suited For	Write Amplification	Read Amplification	Production Systems
B+ Tree	Balanced N-ary Tree (4KB - 16KB pages)	Point lookups, ACID OLTP, low read latency	High ($10\times - 30\times$)	Low ($O(\log_B N)$)	PostgreSQL, SQLite, MySQL InnoDB
LSM Tree	MemTable + Multi-Level SSTables + Bloom Filters	Write-heavy ingestion, high-throughput time-series	Low to Medium ($2\times - 10\times$)	Medium ($O(L \cdot \log N)$)	RocksDB, LevelDB, Cassandra, Pebble
Columnar	Column chunks, dictionary encoding, vector buffers	Analytical OLAP aggregations, range scans	Batched / MergeTree	Minimal for queried columns	ClickHouse, DuckDB, Parquet, Snowflake
Segmented Log	Append-only sequential files + index offsets	Streaming event journals, ordered messaging	Lowest ($1\times$)	Low for sequential consumption	Apache Kafka, Apache Pulsar, Redpanda
In-Memory	SkipList, Adaptive Radix Tree (ART), Hash Table	Sub-millisecond latency, fast cache, session stores	$0$ (RAM), batched AOF/RDB	$O(1) - O(\log N)$ (RAM speeds)	Redis, Memcached, Dragonfly