In my experience, practically every single time I've used absl::btree_map, I've ended up reverting for performance reasons to either a flat hash map or, in some relatively rare cases, a sorted vector map (despite its O(n) insert/erase) because the constant factors are _so doggone low_. The experience remains: btree_map (or SkipLists, or whatever) has (in my experience) essentially never, in over a half million lines of C++, actually remained in the code.
Also, I presume (based on the implementation details, not based on actual use) that roaring bitmaps have some reasonable iteration API that would make them relevant even to the ordered comparison.
The important thing here is that if anyone wants to contend that someone should use their new data structure because it's better or faster or more optimal for some particular use case, it's important for them to demonstrate that they've thoroughly investigated the data structure space around their proposal. Comparison against std::map and std::unordered_map simply doesn't demonstrate the kind of comprehensive knowledge that I would expect from someone who claims that I should use their optimized integer map in my own code.
