And like anything, it depends on what you're doing. If you're doing lots of server-to-server RPCs on big distributed systems like within Google, it's much more reasonable to start with protobuf or the like than with JSON. And not just for performance reasons; having the generated code helps a lot with maintainability. (Some JSON libraries have something similar; some don't.)
* have a shorter exponent and/or mantissa to make each number shorter and the overall series more repetitive. (maybe even skip the mantissa if you only care about order of magnitude! call it the opposite of fixed-point, or just storing the logarithm of the value, whatever terminology works for you...)
* skip the sign bit for a non-negative series (e.g., deltas of a non-decreasing series),
* use a variable-width bitstream encoding like exp-Golomb to represent the exponent and/or mantissa,
* use run-length encoding if it's repetitive (after delta-encoding),
* etc.
One byte per value (or less!) is surely possible in some cases. But I've never done it myself and agree fixed-point sounds much more pleasant. One of my projects represents deltas of durations in units of 90,000ths of a second, varint-encoded. I much prefer that to dealing with float seconds.
edit: also, ~2.5KB of that is due to padding in power_of_ten_components. I wonder if it'd be better to split the uint64_t field into two uint32_ts to avoid this.
[1] https://github.com/lemire/fast_double_parser/blob/master/inc... [2] https://news.ycombinator.com/item?id=22440894 [3] https://en.wikichip.org/wiki/amd/ryzen_embedded/r1606g