Impacts of data consistency levels in cloud-based NoSQL for data-intensive applications

When using database management systems (DBMSs), it is common to distribute instance replicas across multiple locations for disaster recovery and scaling purposes. To efficiently geo-replicate data, it is crucial to ensure the data and its replicas remain consistent with the same and the most up-to-date data. However, DBMSs' inner characteristics and external factors, such as the replication strategy and network latency, can affect system performance when dealing with data replication, especially when the replicas are deployed far apart from the others. Thus, it is essential to comprehend how achieving high data consistency levels in geo-replicated systems can impact systems performance. This work analyzes various data consistency settings for the widely used NoSQL DBMSs, namely MongoDB, Redis, and Cassandra. The analysis is based on real-world experiments in which DBMS nodes are deployed on cloud platforms in different locations, considering single and multiple region deployments. Based on the results of the experiments, we provide a comprehensive analysis regarding the system throughput and response time when executing reading and writing operations, pointing out scenarios where each DBMS could be better employed. Some of our findings include, for instance, that opting for strong data consistency significantly impacts Cassandra's reading operations in the single-region deployment, while MongoDB writing operations are most affected in a multi-region scenario. Additionally, all of these DBMSs exhibit statistically significant variations across all scenarios in the multi-region setup when the data consistency is switched from weak to stronger level.