Files
VietBui_22921300_2016.pdf
Closed access - Adobe PDF
- 1.17 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Standing between the unachievable strong consistency and the undesirable eventual consistency, causal consistency is increasingly considered a valuable model to study. It is because this model ensures causal dependencies between operations to be respected, avoiding unwanted scenarios, such as in a Facebook-like application, a sensitive photo should not be presented to disallowed users before its privacy setting applied. Recently there are several implementations that applied potential causality notion to achieve causal consistency like in COPS and ORBE architectures. However, using potential causality seems to be quite expensive because it creates too many causal relations between objects, introducing a lot of unnecessary overhead metadata as the cost of dependency tracking mechanism. In this thesis, we studied the explicit causality approach, which offers programmers the ability to specify which objects are actually causally related, thus dramatically reduce this overhead. The thesis consists of 2 main parts: (1) a review of this concept’s advantages in comparison to several current potential causality related implementations and (2) our implementation of a client library that enforces explicit causality in one of the geo-distributed noSQL data stores. We conducted several sets of experiments for evaluation, using Yahoo Cloud Service Benchmark and real datasets of several social networking application (Twitter and Reddit) as input, setting Grid’5000’s 2 data centers as the environment. The results show promising figures where our approach outperforms the COPS’ approach in term of reducing metadata overhead burden. This enhancement may encourage further studies about this approach in the future.