It's because it was built with high granularity and unlimited metrics as a key differentiator from the beginning. The core is written in pure C, optimized to death. Even long-term retention was initially sacrificed, in order to be able achieve that high performance, with minimal resource needs.

Long term retention is now possible, but with relatively high memory requirements, depending on how many metrics are collected. Again, it was a decision to never give up realtime granularity and speed, even at the cost of writing our own timeseries db in C and utilizing more memory.