Title: CedrusDB: Persistent Key-Value Store with Memory-Mapped Lazy-Trie

Abstract: Most of today's persistent key-value stores are based either on a write-optimized Log-Structured Merge tree (LSM) or a read-optimized B+-tree, both of which assume that there is not enough memory to cache the entire data set. As RAM is becoming cheaper and more abundant, it is time to revisit the design of persistent key-value storage systems. This paper introduces a new design called "lazy-trie" to index the persistent storage, a variant of the hash-trie data structure, optimized to cache entirely in memory. A lazy-trie achieves near-optimal height and has practical storage overhead, but does not support range queries.
We implemented CedrusDB, a lazy-trie-based persistent key-value store. The implementation uses memory-mapping. The lazy-trie organization in virtual memory allows CedrusDB to better leverage concurrent processing than other organizations. CedrusDB achieves comparable or superior performance for both popular LSM- and B+-tree-based key-value stores in mixed workloads.