You can still do better by not doing any TSC to nanosecond conversion. Instead, you inject clock adjustments into your log stream allowing you to convert when you decode the logs rather than when you generate the logs.

Using the final version, you would just make the cache refresh() function emit the clock adjustment log entry instead of actually caching anything. Then, any later log entry TSC would implicitly be relative to that clock adjustment log entry when you decode the log. Worst case you would need to persist every clock adjustment entry even when sampling, but that would still only be on the order of a few KB/s at worst and you could still drop entrys with no non-clock adjustment entrys between them.

Yep, I should probably have mentioned this option for completeness. In practice, however, it only saves you a few cycles/nanos and adds more complexity and failure points downstream.