HyperDex
HyperDex is a distributed key-value store that uses secondary-attribute coordinates to place objects and route searches across servers.
The 2012 design divides attributes into low-dimensional subspaces. A search uses one subspace selected to minimize contacted servers. The implementation stores a complete object per subspace and adds f+1 replicas per region to tolerate f failures. Index copies and failure replicas are distinct costs. [1]
An immutable key subspace anchors updates. Value-dependent chaining includes both old and new locations when an attribute change moves an object; the new copy is stored before the old one is removed. [1]
Single-key get/put operations are linearizable, but search may return both the committed and concurrently updated versions of a matching object. That is not a single-version snapshot guarantee. Applications must distinguish key-operation consistency from search consistency. [1]
Warp's 2015 design extends HyperDex with acyclic transactions for multi-key serializability. Clients buffer writes and cache reads in a transaction context; the commit protocol validates reads and orders conflicting transactions. These are additional transaction mechanisms, not consequences of the base store's linearizable get/put operations. [2]
Relations
Limitations
- Range predicates require an order-preserving coordinate mapping; arbitrary hashing does not preserve numeric order. [1]
Citations
- Escriva, Wong and Sirer. HyperDex: A Distributed, Searchable Key-Value Store. SIGCOMM 2012, pp. 25-36; sections 2.3, 3.2 and 4.
- Escriva, Wong and Sirer. Warp: Lightweight Multi-Key Transactions for Key-Value Stores. arXiv:1509.07815v1, 2015; sections 2 and 2.1.
Sources & caveats
- Historical source review of the SIGCOMM 2012 design and Warp's 2015 preprint, not a benchmark replication.
- Research classification describes the designs reviewed here, not current maintenance or commercial availability.