With the new linkedin spark connector we should now be able to remove the old `tfrecords` format in favour of `tfrecord` only. This would also allow us to solve #29