Recovering Mobile Users in the Context of Internet Transactions (iTXs)

With the expansion of Web sites to include business functions, a user interfaces with e-businesses through an interactive and multi-step process, which is often time-consuming. A loss of connection, or any other system failure, can result in the loss of work accomplished prior to the disruption. This work must then be repeated upon subsequent reconnection { often at signi cant cost in time and computation. In mobile wireless environments, this \disconnection-reconnection-repeat work" cycle may cause web clients to incur substantial monetary as well as resource (such as battery power) costs. With the introduction of high-speed 3G networks, the use of mobile devices to access Internet applications will no doubt increase dramatically, making this problem more acute. Consider a scenario in which a user is buying an airline ticket over his wireless Internet connection. He rst logs on to the airline site with his frequent ier number, then checks his frequent ier mileage. He then enters his preferred travel dates and destination, and chooses among the itineraries o ered by the site, selects his seat, enters his credit card information, and receives a con rmation of the purchase. If the user's wireless connection drops (as occurs frequently [6]) late in this interaction, e.g., during the purchase step, he must reconnect to the Internet, and restart the sequence of actions. This includes expensive steps, such as processor-intensive login and authentication, I/O intensive database lookups, as well as consume valuable connections to the database. Signi cant amounts of time and e ort are wasted in redoing previously completed work, and, in the case of a wireless environment, there are additional costs in battery resources and airtime. In the scenario described above, we would like to be able to avoid the repetition of work (computation, communication, I/O) required after a connection disruption. Ideally, we would like to reduce the cost of recovering a user's interaction. Much like a database transaction, in many ways, this interaction consists of a number of actions aimed at achieving a particular goal (or set of goals). However, the recovery of such interactions is quite di erent from the classical notion of transaction recovery. In recovering transactions, the focus is on ensuring that the underlying database system is rendered consistent. Thus, if a transaction prematurely aborts, the transaction is rolled back and is resubmitted after the database system recovers. Here, though, the database and application systems have the necessary recovery components in place to recover the database system. These system recovery components do not address the situation where the client portion of a user's interaction history is lost in a connection failure. What is needed here is a mechanism through which users can eÆciently and quickly restart from an appropriate point prior to disconnection (yet subsequent to the beginning of the interaction) so that the interactions can resume from that intermediate point. Thus, our concern is not \system recovery", but rather \user recovery". Note that the \user recovery" problem is not equivalent to the classical database transaction-processing and recovery problem. To illustrate this point, we brie y discuss the di erences between the two problem domains. The theory and application of transaction-based processing in the context of database systems are well-researched topics. Early work, such as [1], describes the basic tenets of transaction processing associated protocols. Long transactions (e.g., Sagas) are discussed in [5]. The metaphor of transaction processing has been extended to other application areas. The ACTA framework [3] describes a formal framework for extending the idea of a transaction to other areas. For example, [2] describes work ow processing using transactions and work ow templates, while [7] utilizes transactions to model mobile interactions. All the works cited above have one idea in common: the activity modeled as a transaction consists of a pre-de ned sequence of operations (typically called a template in the work ow literature). However, user activity with the Web often progresses in a \stream-of-consciousness" fashion, where neither the user nor the site knows the which operation will come next in the interaction.