Scalable transaction management for partially replicated data in cloud computing environments

We present here a scalable protocol for transaction management in key-value based multi-version data storage systems supporting partial replication of data in cloud and cluster computing environments. We consider here systems in which the database is sharded into partitions, a partition is replicated only at a subset of the nodes in the system, and no node contains all partitions. The protocol presented here is based on the Partitioned Causal Snapshot Isolation (PCSI) model and it enhances the scalability of that model. The PCSI protocol is scalable for update transactions which involve updating of only local partitions. However, it faces scalability limitations when transactions update non-local partitions. This limitation stems from the scheme used for obtaining update timestamps for remote partitions, causing vector clocks to grow with the system configuration size. We present here a new protocol based on the notion of sequence number escrow and address the underlying technical problems. Our experimental evaluations show that this protocol scales out almost linearly when workloads involve transactions with remote partition updates. We present here the performance of this protocol for three different workloads with varying mix of transaction characteristics.