Customer issues that came up: * Customer needs to run on a lot of samples which means launching a whole bunch of a-c-r processes. The feature that automatically transfers from HTTP to Keep happens before submitting the workflow. This will scale much better if the data transfer happens on the compute node when the workflow actually launches. * The collection is named "Downloaded from http://..." and if the URL is too long, it will exceed the 255 character limit on collection names. a-c-r needs to account for the limit (probably also including the timestamp that gets added by ensure_unique_name) and trim the name to a valid length so it won't get rejected.