摘要
Apache HBase是当前互联网公司广为使用的开源NoSQL数据库.小米公司几乎所有的在线业务,包括米聊和小米云服务,都使用HBase作后端存储.在海量数据处理应用中,发现HBase的功能存在一定的缺失,在写入吞吐量等性能度量上存在瓶颈.本文针对移动互联网在线业务的特点,基于现有HBase版本提出可保证数据一致性的局部二级索引功能和反向扫描功能,并提出了可配置的细粒度跨集群复制概念,显著提升了单机写吞吐性能.经过全面测试,上述设计和改进方案极大提升了HBase的应用性能,并被部署到了小米的生产集群中.
As an increasingly popular open-source database, Apache HBase has been used by the back-end storage system of Xiaomi inc. for almost all the online services such as MiTalk and Mi Cloud. Facing multivariate requirements of mobile Intemet, HBase has exhibited several drawbacks that cannot fully support Xiaomi＇s business development,such as lacking of secondary index,reverse scan, configurable replication, etc. We therefore developed a series of new features to tackle these problems and moreover, improved the writing performance markedly by introducing a new writing model. All these enhancements have been evaluated through extensive ex- periments before eventually deployed in Xiaomi＇s production clusters.
出处
《小型微型计算机系统》
CSCD
北大核心
2017年第5期1158-1163,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金面上项目(61170022)资助
国家自然科学基金重点项目(91118003)资助
关键词
HBASE
跨集群复制
反向扫描
跨行事务
HBase
cross-cluster replication
reverse scan
cross-row transaction
