Ha, you mean is it memoryless? Yes. You can just ship the whole generation directory around. To roll back generations, just delete newer ones and restart the serving layer. etc. ... View more

The project uses Maven for builds. Eclipse supports Maven builds directly, so therefore supports Eclipse. (Or if there are any problems, it is in the Eclipse integration with Maven.) I use IntelliJ and recommend it, but, there is nothing additional to say here about working with Eclipse. It just works, because the Maven build works, and Eclipse uses Maven builds.   Have you modified the build? You say for example that you had to add the avro plugin to pluginManagement, but it is certainly already there in the parent POM: https://github.com/cloudera/oryx/blob/master/pom.xml#L671   You say there is a compile error, but the dependencies you seem to be missing are correctly specified in the POM already.   That is: you can verify that the build is correct with "mvn compile".   You need to clarify what you are doing because I don't think you are building the project as-is. ... View more

Hard filtering rules need to be implemented in a RescorerProvider, or in logic on the caller side. Tagging users and items with a locale could make sense. It would function as a soft filter nudging people towards things in the same locale. That could be useful as well, but is a different thing from implementing business rules. If your items and users are nearly completely disjoint by locale (e.g. very few items are available in multiple locales and very few users shop in multiple locales) then separate models might be the best way to go. No filtering logic needed although you then manage a model per locale. But the models are smaller and easier to handle. If there is moderate overlap, then a unified model can benefit from the cross-locale learning. ... View more

Yes, that's the same me. Please have a look at the project page for some simple examples to get started: https://github.com/cloudera/oryx There is no Mongo integration per se, but you could always hack on it. ... View more

This is good. There is no performance difference between computing 10 and 100 recommendations since it still considers all non-filtered items each time. (OK I suppose it takes a tiny bit longer to send 100 results over the network than 10.) The results are not precomputed but computed on the fly each time. ... View more

Oops, fixed. Yes I'm using CDH5b1 too, so that's not a difference. Can you compile from HEAD to make sure we're synced up there? you may already be, just checking. I can make a binary too. Any logs would be of interest for sure. I suppose I would suggest trying again with clearly small values for features (like 10) and clearly small values for lambda (like 0.0001) to see if that at least works. I would expect a lower number of features might be appropriate given there are a smallish number of items. You might try the optimizer again with lower ranges for both. More features encourages overfitting and more lambda encourages underfitting, so they kind of counter-act. It's possible you find a better value when both are low. ... View more