I think we should record expected MD5 sums for fetched files in datasets. This way we can give appropriate error messages when a fetch was incomplete, avoiding issues like #8080.