FuseDB

What it is

FuseDB is an embedded KV engine for small hot mutable state, optimized for predictable tail latency on update-heavy workloads.

It is designed for workloads such as rate limiting, counters, sessions, quotas, and other small metadata/state-serving use cases where point reads and point updates dominate.

Important

FuseDB is currently in the research stage. The architecture, APIs, and on-disk format are expected to change.

Why this exists

Motivation

Existing storage engine architectures make different trade-offs:

LSM trees

• often provide strong write throughput
• can suffer from latency instability under compaction pressure
• may rewrite data multiple times across levels
• can accumulate compaction debt that affects tail latency

B+ trees

• provide strong ordered access and efficient range traversal
• may suffer from checkpoint and dirty-page pressure under heavy update workloads
• can experience cache churn and contention on hot paths

FuseDB explores a different design: leaf-local immutable segments with local merge, aiming to reduce global write amplification and improve p99 predictability for small hot-state workloads.

Target workloads

Optimized for

• small keys and values
• point reads
• point updates
• overwrite-heavy workloads
• counters and increments
• hot keys and hot ranges
• bounded or secondary range scans

Not optimized for

• large blobs
• analytical scans
• general-purpose OLTP replacement
• distributed workloads

Warning

FuseDB is not intended to be a general-purpose replacement for RocksDB, B+ tree engines, or full OLTP databases.

Core ideas

• tree-based key-range routing
• mutable in-memory buffer per leaf
• immutable leaf-local segments
• local merge within a leaf instead of global LSM-style compaction
• leaf split on overflow or hot-range pressure
• in-buffer update merge for mergeable mutations
• buffer-to-segment merge for leaf-local materialization
• per-leaf seqno for visibility and merge ordering

Current status

• research stage
• architecture and workload model are being refined
• not production-ready

Roadmap

• Core engine
  • in-memory buffer with update merge
  • immutable leaf segment
  • local merge between buffer and segment
  • basic scheduler
• Leaf split
• WAL and recovery
• First benchmarks
• Tuning, merge policy, and scheduler optimization