During the training phase, we use DataFrame for data processing. To ensure that the feature processing functions are consistent between online prediction and training, we also convert data to DataFrame format during online prediction to reuse the feature processing functions from training. However, we have found that DataFrame processing performance is quite poor.