Distributed Facet Count Recommendations

I ran into the issue of incorrect counts for term facets (has been
discusses previously in this group). I was wondering whether there are
any recommendations how to ensure precise facet counts. I played with
the facet size (using number of shards times desired facet size) but
still did not get precise counts. Is there any support at the search
type side (and if yes, what is the performance implication of such an
approach)?