The size of your data isn't going to be the limiting factor, it's the complexity of the model you will be updating. If it's a simple naive Bayes model then that's pretty easy to implement. If you want something more sophisticated e.g. multi-connected network then the model complexity will determine if you can do exact inference or if it will require trade-offs e.g. approximate algorithms.
Kevin Murphy has a recently updated comparison of software for Bayesian inference at http://www.cs.ubc.ca/~murphyk/Software/bnsoft.html. One open source package I'm exploring is libDAI (http://cs.ru.nl/~jorism/libDAI/) which is in C++ but I assume it's callable from Java. It supports multiple inference methods, including loopy belief propagation which seems to be a pretty good approximation algorithm.