This paper investigates the problem of network partitions in Dependable Distributed Learning Objects Repositories. The general definition of dependability is the ability to deliver service that can justifiably be trusted. This is an integrating concept that has an important attribute: the availability, which deals with readiness for usage. Replication is used to enhance availability and performance in distributed systems. Replication maintains replicas (copies) of critical data on multiple computers and allows access to any of them.

Replication in the data centric applications requires state synchronization in order to maintain the system correctness. Replica consistency defines the correctness of replicated data (how replicas of the same logical object may differ from each other) and constraint consistency defines the correctness of the system state with respect to the set of data integrity rules. If consistency needs to be ensured all the time, such systems soon become partially unavailable if node and link failures occur. The basic problem of network partitions due to these failures is that there is no way of knowing what is happening in the other parts of the system. The trade-off between availability and the system consistency can be configured in such systems. It is possible to deliver some kind of services in a partitioned system acting optimistically. This means to promote availability accepting some kind of operations provisionally and undoing them at a later stage, if necessary.

We aim at a concept which optimizes dependability in Distributed Learning Objects Repositories. In the eLearning, most of the existing approaches only tolerate node failures and make the unrealistic assumption that the network cannot be partitioned. It is possible to model complex learning structures in which educational activities are developed by different users at the same time working collaboratively. Also it is possible to define sequences of educational activities and the conditions in which each activity should be done. There are different standards that specify how to do that. Nevertheless, none of them includes elements supporting the availability of the learning objects.

The contributions of this paper are twofold. First, it is explained how the high availability property can be satisfied by the learning objects repositories, which is not included in the current eLearning standards. Second, an optimistic partition-processing approach is proposed, which can make learning objects available when network partitioning occurs. Novel constraint taxonomy is included to be used when the learning units are designed. To gain high availability the Learning Objects Repository is replicated over different networked computers. However, this approach does not guarantee that the same learning object in different partitions can maintain a consistent value. This availability gain is traded against consistency since several replicas of the same objects could be updated separately.

A mechanism that supports continuous service in presence of multiple partitions is also defined. Once partition terminates, divergences in the replicated state need to be reconciled. One way to reconcile the state consists in letting the application manually solve inconsistencies.