about MSA sampling for self-distillation samples

From 1.3:
> At training time, extra augmentation is added to distillation dataset examples by uniformly sampling the MSA to 1000 sequences without replacement (this is on top of any sampling that happens in the data pipeline, see subsection 1.2).

And there is an "MSA block deletion" (1.2.6, Algorithm 1).

I want to know, is "sampling the MSA to 1000 sequences" before "MSA block deletion" or after?

Another question is, is "sampling the MSA to 1000 sequences" / "MSA block deletion" is called at each recycling iteration independently, or is called at the beginning of the data pipeline one time?
In other words, do we place it at the beginning of `nonensembled_map_fns` or the beginning of `ensembled_map_fns` ?

https://github.com/deepmind/alphafold/blob/9c4ac8a92125942f73813649d9f6885532c1ee97/alphafold/model/tf/input_pipeline.py#L33-L122