XGBoost optimized computation in several respects to enable multithreading by keeping data in memory in compressed column blocks, where each column is sorted by the corresponding feature value. XGBoost computes this input data layout once before training and reuses it throughout to amortize the additional up-front cost. The search for split statistics over columns becomes a linear scan when using quantiles that can be done in parallel with easy support for column subsampling.

The subsequently released LightGBM and CatBoost libraries built on these innovations, and LightGBM further accelerated training through optimized threading and reduced memory usage. Because of their open source nature, libraries have tended to converge over time.

XGBoost also supports monotonicity constraints. These constraints ensure that the values for a given feature are only positively or negatively related to the outcome over its entire range. They are useful to incorporate external assumptions about the model that are known to be true.