Replex
Replex is a research datastore that partitions complete data copies by different index keys so replication also supplies query indexes.
A replex contains every row, sharded and sorted for one index. Hybrid replexes also store full rows, not erasure-coded fragments; their placement limits which partitions must be searched after a failure. [1]
For p = n1 * n2 partitions, equation 2 maps a row to n2 * (fA mod n1) + (fB mod n2). With the hybrid layer intact, an unavailable A partition has n2 possible hybrid targets; an unavailable B partition has n1. These are target counts, not latency multipliers. [1]
Setting (n1,n2) = (1,p) reduces the equation to B placement. Section 3.2's extreme-case prose instead names A; section 5.3 agrees with the equation. The linked exhaustive residue model makes the distinction explicit. [1] [2]
Committed updates can become visible through one index before another while completion is still in flight. Writes involving a failed node wait for recovery. These protocol limits are separate from read fan-out. [1]
Relations
- related
- Chain replication, vCorfu
Citations
- Tai et al. Replex: A Scalable, Highly Available Multi-Index Data Store. USENIX ATC 2016, pp. 337-350; sections 2.2-3.2 and 5.3.
- Hesela: exhaustive synthetic Replex placement enumeration and parameter orientation.
Further reading
Sources & caveats
- Source review plus an exact placement calculation, not a prototype replication or recovery benchmark.
- The 2016 prototype modified HyperDex with its internal fault-tolerance parameter set to zero. The exact evaluated code revision has not been established. [1]