Biblio

Traditional cloud stacks are designed to tolerate random, small-scale failures, and can successfully deliver highly-available cloud services and interactive services to end users. However, they fail to survive large-scale disruptions that are caused by major power outage, cyber-attack, or region/zone failures. Such changes trigger cascading failures and significant service outages. We propose to understand the reasons for these failures, and create reliable data services that can efficiently and robustly tolerate such large-scale resource changes. We believe cloud services will need to survive frequent, large dynamic resource changes in the future to be highly available. (1) Significant new challenges to cloud reliability are emerging, including cyber-attacks, power/network outages, and so on. For example, human error disrupted Amazon S3 service on 02/28/17 [2]. Recently hackers are even attacking electric utilities, which may lead to more outages [3, 6]. (2) Increased attention on resource cost optimization will increase usage dynamism, such as Amazon Spot Instances [1]. (3) Availability focused cloud applications will increasingly practice continuous testing to ensure they have no hidden source of catastrophic failure. For example, Netflix Simian Army can simulate the outages of individual servers, and even an entire AWS region [4]. (4) Cloud applications with dynamic flexibility will reap numerous benefits, such as flexible deployments, managing cost arbitrage and reliability arbitrage across cloud provides and datacenters, etc. Using Apache Cassandra [5] as the model system, we characterize its failure behavior under dynamic datacenter-scale resource changes. Each datacenter is volatile and randomly shut down with a given duty factor. We simulate read-only workload on a quorum-based system deployed across multiple datacenters, varying (1) system scale, (2) the fraction of volatile datacenters, and (3) the duty factor of volatile datacenters. We explore the space of various configurations, including replication factors and consistency levels, and measure the service availability (% of succeeded requests) and replication overhead (number of total replicas). Our results show that, in a volatile resource environment, the current replication and quorum protocols in Cassandra-like systems cannot high availability and consistency with low replication overhead. Our contributions include: (1) Detailed characterization of failures under dynamic datacenter-scale resource changes, showing that the exiting protocols in quorum-based systems cannot achieve high availability and consistency with low replication cost. (2) Study of the best achieve-able availability of data service in dynamic datacenter-scale resource environment.