| Summary: | Databases are a fundamental component of any web service, storing and managing all the service data. In large-scale web services, it is essential that the data storage systems used consider techniques such as partial replication, geo-replication, and weaker consistency models so that the expectations of these systems regarding availability and latency can be met as best as possible. In this dissertation, we address the problem of executing transactions on data that is partially replicated. In this sense, we adopt the transactional causal consistency semantics, the consistency model where a transaction accesses a causally consistent snapshot of the database. However, implementing this consistency model in a partially replicated setting raises several challenges regarding handling transactions that access data items replicated in different nodes. Our work aims to design and implement a novel algorithm for executing transactions over geo-partitioned data with transactional causal consistency semantics. We discuss the problems and design choices for executing transactions over partially replicated data and present a design to implement the proposed algorithm by extending a weakly consistent geo-replicated key-value store with partial replication, adding support for executing transactions involving geo-partitioned data items. In this context, we also addressed the problem of deciding the best strategy for searching data in replicas that hold only a part of the total data of a service and where the state of each replica might diverge. We evaluate our solution using microbenchmarks based on the TPC-H database. Our results show that the overhead of the system is low for the expected scenario of a low ratio of remote transactions.
|
|---|