[08:42:10] <mgerlach>	 chrisalbon: here is some info on stat1004-8 https://wikitech.wikimedia.org/wiki/Analytics/Systems/Clients
[14:21:34] <chrisalbon>	 oh awesome, thanks!
[18:53:28] <chrisalbon>	 So I misspoke in our ML/Research sync. One of the questions brought up was if it would be possible to preprocess an API request before requesting a prediction of Lift Wing's model. For example: take a diff, convert it into features, request a prediction from a model, then return that prediction. 
[18:53:28] <chrisalbon>	 I said that wouldn't be in the MVP, but I actually think we will be able to do it easily using KFServing's transformer system.
[20:16:53] <isaacj>	 chrisalbon: great news! I haven't worked with KFServing but hopefully meets our needs. the general point was that the input to our models is almost always a revision ID or article title or some other ID that we then use to gather data via the APIs. it's very rare that the input is e.g., a string of text that we want processed. the ORES text-based topic classification is a good example of this. there are several steps: revision ID -> 
[20:16:53] <isaacj>	 revision text (via API) -> basic text processing -> sequence of word vectors (via embedding lookup) -> document embedding (average the word vectors) -> input into model for prediction -> topic scores. if these steps happen at different systems, then it feels like potentially a lot of IO to pass the preprocessed article text or just odd that the embeddings are maintained in a separate system from the model (because the model is 
[20:16:53] <isaacj>	 trained with a specific set of embeddings).
[20:23:11] <chrisalbon>	 yep all that sounds reasonable. Hopefully this will keep that load light for you all 
[20:31:46] <isaacj>	 excellent - thanks! don't hesitate to let us know if we can provide any more feedback or test out anything etc. too