The prediction coming out of XGBoost has to go through Python again, serialized as JSON, and sent via HTTP to the client. XGBoost itself is written in C++, and it's Python library only provides a convenient interface. Since the features are externally stored, they need to be passed through Python and into XGBoost. In Python, most of the bottleneck comes from having to fetch and deserialize Redis data. In this benchmark, PostgresML outperformed Python and Redis, running on the same machine, by a factor of 8. Throughput is defined as the number of XGBoost predictions the architecture can serve per second.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |