For rank:pairewise objective, it needs partition the dataset on the group by user, and set numWorkers to be same as the partitions

We are training a pairwisde model with rank:pairewise objective. We found that we have to pre partition our dataframe on group column when we setting the numWorkers > 1, otherwise the model performance will degrade. And we also found that we have to set the numWorkers to be same as the pre partition number, otherwise the model performance will degrade too.