From my experience, it was pretty speedy even on some (relatively) large datasets. The decoupling of compute and storage makes it appealing compared to Redshift if you have bursty load on your cluster.