A simple totally ordered broadcast protocol

Asimpletotallyorderedbroadcastprotocol Benjamin Reed Flavio P. Junqueira Yahoo! Research Yahoo! Research Santa Clara, CA - USA Barcelona, Catalunya - Spain breed@yahoo-inc.com fpj@yahoo-inc.com ABSTRACT chines providing the service and always has a consistent view of the ZooKeeper state. The service tolerates up to f crashThis is a short overview of a totally ordered broadcast pro- failures, and it requires at least 2f + 1 servers.tocol used by ZooKeeper, called Zab. It is conceptually Applications use ZooKeeper extensively and have tenseasy to understand, is easy to implement, and gives high to thousands of clients accessing it concurrently, so we re-performance. In this paper we present the requirements quire high throughput. We have designed ZooKeeper forZooKeeper makes on Zab, we show how the protocol is used, workloads with ratios of read to write operations that areand we give an overview of how the protocol works. higher than 2:1; however, we have found that ZooKeeper’s high write throughput allows it to be used for some write 1. INTRODUCTION dominant workloads as well. ZooKeeper provides high read throughput by servicing the reads from the local replica ofAt Yahoo! we have developed a high-performance highly- the ZooKeeper state at each server. As a consequence, bothavailable coordination service called ZooKeeper [9] that al- fault tolerance and read throughput scales by adding serverslows large scale applications to perform coordination tasks to the service.