datasets generated for dependent job launches should be more restrictive.
We are getting situations, usually when a user cancels a submission after some files have been generated, that the dependent stage of the next submission get both files from this launch and from the cancelled one, leading to duplication.
To fix this, we should add a check on creation time on the dataset definition, that it is later than the start of the submission. This will eliminate the problem, as long as files get their metadata declared in the job. If files are being declared later, say by an FTS, then we could still get some overlap here, but it does give us a recipe to give users to avoid it entirely.