The Design And Implementation Of Metadata Service High Availability For Distributed File System

The traditional stand-alone file system cannot meet the explosive growth of data storage needs in today's era of massive data.The distributed file system has been widely used in society because of its large capacity,high performance,and high scalability.However,how to ensure the availability of the system is very important due to the important positioning of the distributed file system.At present,most mainstream distributed file systems use metadata and actual data to be stored separately.Its operation ratio on metadata is far greater than that on actual data.Therefore,the availability of metadata service largely determines the availability of the whole distributed file system.AFS,a distributed file storage system independently developed by an Internet company in China,adopts centralized management of metadata.At present,the high availability scheme of metadata service adopts master-slave mode,which is similar to HDFS hot backup node to realize high availability of service.However,this scheme relies more on shared storage.It is prone to a single point of failure of switches because its shared storage system relies on a keepalive mechanism.In particular cases,the host cannot be switched,which is not conducive to the promotion of the AFS file system.This thesis upgrades and improves the metadata service high availability scheme of the AFS file system based on the research of metadata service high availability technology.The main work is as follows:1.This thesis analyzes the problems existing in the existing metadata service high availability scheme of AFS and proposes to realize log data synchronization between nodes based on Raft protocol,to realize the high availability of metadata service.2.This thesis analyzes the key links of Raft protocol in detail,including the concrete implementation of master selection,log synchronization,log compression,and so on.At the same time,a Snapshot scheme based on Learner role is proposed because the business scenario of AFS file system application requires extremely high system availability,in which a single Learner node does not participate in the election work,but is only responsible for learning the log and completing the Snapshot,thus solving the problem that the service is unavailable during the node's Snapshot.3.This thesis implements the highly available module of AFS file system metadata service in detail based on the new Raft architecture,including but not limited to the workflow of Learner role,log double writing,Snapshot generation,and loading process,operation,and maintenance tools,etc.4.In this thesis,the usability test and performance test of the highly available module under the new Raft architecture are carried out.The results show that under the new high availability scheme,the system service can be restored in a very short time after the system fails,thus ensuring the high availability of system;Moreover,it has little influence on metadata performance and the whole system,and the performance is greatly improved after adopting ssd scheme.At present,this high availability scheme has been applied in company A's AFS clusters,and dozens of clusters have been upgraded online,running well.The new architecture solves the problems existing in the previous old architecture and ensures the availability of system metadata services.