We are facing a similar problem using the _cache_key wherein we see inconsistent results while doing a simple terms lookup. We are currently using elasticsearch 1.6.3 and have a three node cluster setup.
Strangely, the output varies on every third query we make to the cluster e.g. we consistently receive 0, 0, 2, 0, 0, 2... as the doc count and it appears that the cached value is not being synchronized across all the nodes. The logs within each node do not indicate anything untoward either.
I see this was raised in 2015. @pmichel Were you able to find out the reason for the inconsistency?
Has anyone else faced such a problem?