We are still using the nearest neighbor approach as default. This method will be tested along with Kmeans via Nimbus. From preliminary perf. testing, it's also more performant. Differential Revision: https://phabricator.services.mozilla.com/D246551