A study of online construction of fragment replicas

Abstract

High availability in database systems is achieved using data replication and online repair. On a system containing 2 replicas of each fragment, the loss of a fragment replica due to a node crash makes the system more vulnerable. In such a situation, only one replica of the fragments contained in the crashed node will be available until a new replica is generated. In this study we have investigated different methods of regenerating a new fragment replica that is up to date with the transactions that have happened during the process of regenerating it. The objective is to determine which method performs the best in terms of completion time at each of the nodes involved, in different conditions. We have investigated three different methods for sending the data from the node containing the primary fragment replica to the node being repaired, and one method for catching-up with the transactions executed at the node containing the primary fragment replica during the repair process. These methods assume that the access method used by the DB system is B-trees. The methods differ by the volume of data sent over the network, and by the work (and time) needed to prepare the data prior to sending. They consist respectively in sending the entire B-tree, sending the leaves of the B-tree only, and sending the data only; the latter has two alternatives on the node being repaired, depending on whether the data is being inserted into a new B-tree, or whether the B-tree is being regenerated from the leaf-level and up. This study shows that the choice of recovery method should be made considering the network configuration that will be used. For common network configurations like 100Mbits or lower, it is interesting to use methods that minimize the volume of data transfered. For higher network bandwidth, it is more important to minimize the amount of work done at the nodes.