In my opinion, tensorflow transform is likely not the right tool for this job. What would clarify it is if the two datasets are two independent streams that come in at inference time, and need to be merged somehow, an also need to be done in the tensor graph. The last part is the reason I'm saying transform is likely not the right tool, because a group by is not a tensor operation. Tensorflow transform encodes tensor operations into a tensor graph similar to the fit method in sklearn.preprocessing. Tensorflow transform is for things that require a full pass over the training dataset (e.g. compute and store mean, variance etc). Hope this clarifies things, let me know if you need further help with this.