Large-scale learning with infinite of number of classes

I have a learning problem that is somewhat like the Netflix problem (but not exactly). I'll describe the model briefly below.

tl;dr It's a classification problem but with infinitely many classes. Can I use vowpal wabbit for this or some other toolkit that supports large-scale learning on hadoop? (I have hundreds of thousands of examples and hundreds of millions of features.)

Here is the model, which, I think, is pretty standard:

I have N observations of pairs (person, movie). For example:

Count | UserName | WatchedMovie ("class")
------+----------+-----------------------
3     | adamG    | Catch Me If You Can
1     | adamG    | Pulp Fiction
4     | bettyX   | Savages
9     | bettyX   | Schindler's List
2     | bettyX   | A.I.
...

Now I define a logistic (i.e., maximum entropy) model p(m|u) based on the features of users and movies (u=UserName, m=WatchedMovie):

p(m|u) = 1/Z(u) score(u,m)
score(u,m) = \sum_k weight_k feature_k(u, m)
Z(u) = \sum_{m'} score(u,m')

Example features:

feature_1(u,m)=studied_at_harvard(u) && is_steven_spielberg_movie(m)
feature_2(u,m)=studied_at_harvard(u) && has_actor_john_travolta(m)
feature_3(u,m)=lives_in_texas(u) && is_action_movie(m)
...

Each feature has a feature weight, and I'd like to learn the feature weight vector that maximizes p(m|u) over all observed data points (using fast online learning, e.g., gradient descent).

Then, given a learned feature vector, I can make predictions on new users. For example, p(Shrek | christieH) could be high if christieH has similar features to the observed bettyX and Shrek has the same director as Schindler's List which was observed with bettyX, etc.

Summary. Again, the question is: Can I use vowpal wabbit for this, or some other toolkit that supports large-scale learning on hadoop? (I have hundreds of thousands of examples and hundreds of millions of features.)

For vowpal wobbit, all examples I've seen are multi-class and assume that we have few classes only. But in my example I have an infinite number of classes, so I'm not sure how I'd enter the input data for learning.

Some alternative model would be okay too (simple linear instead of log-linear), I just want to do large-scale learning and get a meaningful prediction score for new users and movies based on their features.