I should clean up the documentation, but results_all() is a convenience wrapper that is not too smart in that it simply tries to return all results in one trip to the server. This can fail for multiple reasons related to the size of result sets. The better approach (and the only one in the case of large results sets) is to page through the results:
Unfortunately, the size parameter really requires trial-and-error to find the largest "working" setting since the results can vary quite significantly in volume. Instead, I usually just choose a smallish number like 50 or so and wait a few extra seconds. These calls can, in theory, be parallelized using something like BiocParallel to get really fancy (and introduce complexity).