Rollback Recovery For WebServices

Abstract

Rollback recovery is one of many critical tasks in the design of web services systems. It has become increasingly important as a result of the high load of requests to be processed by the system. It has also been realized that loss of messages do affect the business due to a variety of service failures. Deterministic and non-deterministic event failures are often in web services. Deterministic event failures occurs when processing the messages at service level. But nondeterministic event failures occur mainly due to service inactive state, which can not process the requests. Suitable rollback-recovery techniques are identified previously for message passing systems which are checkpoint-based and log-based techniques. Survey done on message passing systems for non-deterministic event failures concludes that we can recreate the lost state of the service by replaying the saved messages. Survey further points that, we can use both Rollback-recovery techniques to store the non-deterministic event failures in a message passing system. Both mechanisms have their own advantages and disadvantages. The overall goal of this research effort was to identify a suitable rollback-recovery technique for non-deterministic event failures which occur in web services and implement a proof of concept design and implementation to validate the results. To achieve this goal, a detailed study was carried out to understand and decide the more suitable rollback- recovery technique. Log-based recovery mechanism was selected as suitable for web services and using that technique a proof of concept design and Java-based implementation was done for Apache Axis2 web service platform. Implementation follows the Axis2 architectural design mainly the handler chain design which can be modified according to the system design. Tests were carried to validate the results of the design for different web services and different type of services like SOAP and REST. In failure case test, we sent non-deterministic messages and made them to fail. We were able to log those messages, successfully replay them and we were able to produce correct response for those failed requests. In success case test,our system was able to log the all non-deterministic messages and finally it didn't replay any messages because all messages were successful requests. ii Several design variables including the system performance, high I/O performance for database, garbage collection mechanism were identified.