Dependability as a cloud service - a modular approach

Failures of services on cloud platforms are only to be expected. To deal with such failures, one is naturally inclined to use the traditional measure of replication. However, replication of services on distributed cloud platforms poses several challenges that are not well met by today's Java middleware systems. These challenges are the need to isolate state in the application components so that easy migration and recovery are possible and the requirement for client transparency when dealing with different replicated service instances. For example, Java Enterprise Edition (JEE) makes it difficult to have transparent replication of services due to the above two reasons plus the fine-grained nature of interactions between its components (the Enterprise Java Beans). In this paper, we show parts of the design of OSGi, a specification defining a dynamic component system in Java, that make it suitable for the above task. We then propose two extensions to OSGi which will allow exposing and exporting application component state and transparent invocation of service instances. These two together can enable easy replication and recovery from failures in cloud environments. We show through experiments that our prototype can migrate a failed service quickly enough to a new machine so that a client experiences only a moderate increase in service invocation time during system recovery.