In general, when dealing with buckets, counts must be split over le. But one metric, hubResponseCodes, sums over code only, which ends up counting each request once per bucket they match, since buckets are "less than or equal" and not mutually exclusive, resulting in over-counting requests by ~15x (the number of buckets). I became suspicious when my deployment had ~3k steady 200 responses every 2 minutes with ~20 active users. The real number was ~180.

sum of buckets over code (before this PR):

Sum of buckets over code, le, showing double count:

Sum of count over code (this PR), matches the le=+inf value above: