COLE: A Column-based Learned Storage for Blockchain Systems
summary
COLE implements a column-based learned index that stores each state's historical values contiguously and replaces Merkle Patricia Trie nodes with epsilon-bounded piecewise linear models plus compact Merkle trees. It uses an LSM-style layout with in-memory B+tree runs, disk runs holding value, index and M-tree files, streaming model construction, and asynchronous merges. Evaluation on Ethereum-like workloads reports reduces storage size by up to 94%, 1.4×–5.4× throughput improvements, and sublinear provenance-proof growth compared to MPT.
AI-written summary · about
chapters (8)
- 0:00COLE and blockchain storage
- 3:28MPT limitations
- 7:03Naive learned index fails
- 8:10COLE’s column-based design
- 13:30Learned models and M-tree
- 15:56Query processing in COLE
- 17:00Experimental evaluation
- 20:17Summary and future
related talks
15mLVMT: An Efficient Authenticated Storage for Blockchain
25mNew Trends In Storing Large Data Silos With Python
20mSlimArchive: A Lightweight Architecture for Ethereum Archive Nodes
45mClickHouse: what is behind the fastest columnar database
46mColumnar Store for PostgreSQL
36mA journey toward the columnar data store - lessons from past discussions and new proposals
38mLight and Adaptive Indexing for Immutable Databases
18mPreview: KV-Stores and Databases