Building social networking services systems using the relational shared-nothing parallel DBMS

摘要

We propose methods to enable the relational model to meet scalability and functionality needs of a large-scale social networking services (SNS) system. NewSQL has emerged recently indicating that shared-nothing parallel relational DBMSs can be used to guarantee the ACID properties of transactions while keeping the high scalability of NoSQL. Leading commercial SNS systems, however, rely on a graph - not relational - data model with key-value storage and, for certain operations, suffer overhead of unnecessarily accessing multiple system nodes. Exploiting higher semantics with the relational data model could be the remedy. The solution we offer aims to perform a transaction as a set of independent local transactions whenever possible based on the conceptual semantics of the SNS database schema. First, it hierarchically clusters entities that are sitting on a path of frequently navigated one-to-many relationships, thereby avoiding inter-node joins. Second, when a multi-node delete transaction is performed over many-to-many relationships, it defers deletion of related references until they are accessed later, thereby amortizing the cost of multi-node updates. These solutions have been implemented in Odysseus/SNS - an SNS system using a shared nothing parallel DBMS. Performance evaluation using synthetic workload that reflects the real SNS workload demonstrates significant improvement in processing time. We also note that our work is the first to present the entity-relationship schema and its relational representation of the SNS database.