Getting consistent scoring best practices

[usteidj]

I've had the case that's described in the "Scores are not reproducible" section of this: Getting consistent scoring.
I am using the ElasticSearch Service in AWS.
I'm going to have a large index (millions of documents), but right now, I only have about 16K.
The pipeline replaces existing documents by doing a delete_by_query and a refresh before adding replacements (some would be overwritten, some would be deleted and some would be added). Along with doing the delete_by_query and refresh, I've got "gc_deletes" : 0 set in the index.
I don't know how often it will happen, but I've already seen one occurrence where I run the exact same search multiple times and I either get one result set where the top score is one thing (and the documents are ordered according to it) or another thing (and documents are ordered likewise).
Now, I don't really have the opportunity to place a "session" identifier in the preference value - I could always place the same value (something like "abc123) in that field, but to me that seems like it might be a problem because every search that will come in will have that exact same identifier. Is that indeed a problem?
Would using the dfs_query_then _fetch handle this problem? There's a note I saw in one of the ElasticSearch documentation articles that said that it shouldn't be used in production.
Finally, is there a strategy to prevent the root cause from presenting itself in the first place, or at least greatly reduce the likelihood? I was wondering whether I need to run a _forcemerge on a regular basis to make sure all deletes are not taken into consideration when scoring happens?
Thanks for any guidance.