Many data-intensive applications and services in the cloud are geo-distributed and rely on geo-replication. Traditional synchronous replication that ensures strong consistency exposes these systems to the bottleneck of wide areas network latencies that affect their performance, availability and the monetary cost of running in the cloud. In this context, several weaker consistency models were introduced to hide such effects. However, these solutions may tolerate far too much stale data to be read. In this PhD research, we focus on the investigation of better and efficient ways to manage consistency. We propose self-adaptive methods that tune consistency levels at runtime in order to achieve better performance, availability and reduce the monetary cost without violating the consistency requirements of the application. Furthermore, we introduce a behavior modeling method that automatically analyzes the application and learns its consistency requirements. The set of experimental evaluations on Grid'5000 and Amazon EC2 cloud platforms show the effectiveness of the proposed approaches.