For many recommender systems users and items are 'embedded' within a latent vector space. The closer the user and the item are in this space, the higher the probability for the item to interest the user. In this talk, I'll describe the pain points and solutions for using embeddings in a high-throughput search based recommender system . From coping with cold start users and items that require on-the-fly embedding generators in serving, to keeping up with the dynamic nature of embeddings that change rapidly with every chunk of new supervised data. We'll cover batch indexing and incremental indexing setups as well as a solution for parallel AB testing of multiple embedding schemes.