Previously, this would pipe tar's output to curl. But curl would buffer the entire output to memory in order to calculate the content-length header. This pointlessly made large tar backups impossible as the backend does not require a content-length header sent with the data.