Learning-to-rank feature assembly

Design the serving path for a learning-to-rank (LTR) model in product search: a query retrieves up to 2,000 candidate items, and for each you must assemble a ~150-dimensional feature vector (query-item features, item features, user-item features), score all 2,000 with a gradient-boosted tree (or small DNN), and return the top 24, all in under 120ms p99 at 30k searches/sec. Focus on feature assembly and consistency, not on training the model. How do you avoid the feature vector being slow or different from what training saw?