There is an increasing need for tools that facilitate business
decisionmaking in the face of uncertain data. The problem of data
uncertainty is becoming acute, due to data integration, automated
information extraction, data anonymization for privacy protection, and the
growing importance of RFID and sensor data. Recently, in joint work between
IBM Research and University of Florida, the MCDB relational database system
has been developed for managing uncertain data, based on a Monte Carlo
approach. This system can handle complicated real-world queries and data,
and has an extensible and flexible uncertainty model, encapsulated via
user-defined "value generation" (VG) functions. MCDB also allows
sophisticated, data-intensive stochastic modeling and prediction to be
performed close to the data. The key technical idea is to process a query
plan once, but over "tuple bundles" that encapsulate possible worlds, rather
than ordinary tuples; pseudorandom number seeds are used to compress such
bundles whenever possible, in order to provide acceptable performance. We give an
overview of the MCDB system, and also describe a recent effort to implement
the MCDB functionality within the the setting of cloud computing,
specifically, Hadoop, augmented with the Jaql query language. We focus on
cloud computing not only because it is an increasingly popular and
ubiquitous computing model, but also because it can potentially speed up the
CPU-intensive MCDB computations via massive parallelism, and permits
straightforward extension of MCDB functionality to certain types of
non-relational data. Key challenges in this setting are to manage the
pseudorandom number seeds that form the basis of the Monte Carlo
computations, and to understand the tradeoffs between "inter-tuple"
parallelism and "intra-tuple parallelism" (generating possible worlds for a
single tuple in parallel). We also briefly describe our ongoing efforts to
develop a realistic end-to-end business scenario, based on current work by
IBM's AVATAR project to develop principled algorithms for assigning
probabilities to annotated data.