Currently, we're looping over all the requests in the bulk request for each failed lazy rollover (here, and once #108108 is merged, we'll do it in BulkOperation too). While this is probably around O(n) in practice, it's technically O(n^2), if each document was destined to a different failure store and each rollover failed.

A better alternative would be to collect all the failed rollovers and traverse the documents one time.