AB test design for monolith

[uchiiii]

Thank you for sharing great codebase for realtime online learning framework.

I am trying it out. I am curious to know how you design AB test using Monolith. Your paper suggests that features used in prediction are used for training as well, which means that you can't use it for training if you do not predict it. Let's say you conduct an AB test and allocate 5% of traffic to C group and T group each. If you just apply Monolith, a model for C group consumes 95% (= 5% + 60%) of traffic and the other model for T does 5% and this is very unfair. What do you think is the best practice to do AB test fairly. Did you mirror prediction requests to all models?

Thank you for your reply in advance.

[ZackMX]

how did u able to run the native training? could u help me?

[uchiiii]

@ZackMX
Sorry I am not very familiar with it. I just finished reading the paper right now