I am trying to implement efficient paging through all results of a search query using search_after with _doc index ordering on a 12-node cluster with 1 shard that has 11 replicas (elasticsearch 5.6). I'm running into a problem where the "_doc" value returned in the "sort" field of the response is different for the same document across different replica shards. This makes it impossible to page through results consistently.
Is this behavior expected? Do we expect index ordering scores to be different across replicas?