Currently, we are using the same heuristics as for float samples, which mostly means that we put up to 120 samples into one chunk. However, this might create very large chunks for histograms with many buckets. So we should probably stop earlier in that case. On the other hand, for a rarely changing histogram and/or one with very few buckets, the trade-offs might be different compared to floats, and storing more than 120 histograms in one chunk might be more efficient.

This could possibly done later as a non-breaking optimization, but just in case it needs a breaking change in the storage format, let's call it P2. (Also, it seems quite problematic to declare native histograms as stable if we don't know yet about the impact here.)