When an index is built in a rolling fashion (by taking nodes out of a replica set one by one and building the index on them) as described in Build Indexes on Replica Sets, an oplog entry never gets created and the index doesn't get created on the backup. There is currently no safe workaround to create such indexes on backup and the only available solution is a resync of backup which could be very long.

Could we have a way to create an entry in oplog for indexes that already exist and have secondaries (including backup) only create that index if it's not already there?