Use-case dependent. For some analytics, you really want to see the tail ends (e.g. rare events) which sampling can sometimes omit or under-represent.