Landing transactional support on the cloud

Abstract

We took the Master thesis of I. Arrieta-Salinas and M. Louis Rodríguez as a starting point for this project. We are going to deploy a distributed database to be used in a cloud environment as a specific case of Platform-as-a-service. We assume that data is partitioned and several replicas store a copy of a given partition. The clients issue transactions by means of a standard library such as JDBC. To do so, they need information about the data placement that is managed by a Metadata Manager. The Metadata Manager manages the partitioning and the replica placement among all replicas building a replica cluster on each partition. The replication cluster has a few replicas running a replication protocol to provide strong consistency and the rest receive the propagation of updates in a lazy manner. These replicas are logically constituted as onion layers around the core replicas running a given replication protocol. The implementation of this system had several drawbacks that we try to fix in this work. First of all, clients an the MM need to be physically in the same machine which leads to a penalty performance in heavily loaded scenarios. The system was optimized for YCSB that consisted in transactions with a single operation and they are run over two replication protocols: primary copy and active replication that are known to perform badly update intensive scenarios. Moreover, there was no load balancing at all according to replica performance, it was merely a round-robin policy among all replicas at the core level. We try to argument the system limitations (described in more detail in Section 2.1) and to going into the system implementation. This is going to be explained in the rest of this work. The main goals of this project are focused in the different parts of the system. In regard to the Client Module, originally the client was the OLPT-Benchmark, a module that consist in send specific types of transactions to the system by a JDBC connection. In the actual version this module has been modified allowing to the transaction to have more than one operation and several parameters has been introduced to the transaction which allow to the system to treat them differently. Respecting to the Metadata Manager one of the main goals between the others developed in this project is the decentralization of the Client and Meta- data Manager modules physically. The rest of modifications are the creation of a structure that allow to the Metadata Manager to know the architecture of the Replicas Cluster and the development of a new ReplicaChooser function based on the CPU charge allowing a correct load balancing. And finally in the Replicas Cluster has been implemented new protocols that have permitted to run different replication protocols in different partitions simultaneously without the knowledge of the Client and the Metadata Manager.