We're likely to normalize on milliseconds, but at the same time, the implication of floating point arithmetic on all our μs numbers isn't ideal.
Our problems were caused by a mix of serialization format (JSON numbers) and not always converting into the language's date/time types at the boundary (sometimes raw epoch seconds/millis were passed around layers of code and only parsed into a date for display. That created opportunities for misinterpretation at every function call.
My general rules for non-performance critical code are
1. Always Parse into a first-class date/time/duration type at the serialization boundary.
2. Always use an unambiguous format (e.g. ISO-8601) for serialization
It's not the most efficient but lets you rely on the type system for everything in your code and only deal with conversion at one place.
For a similar issue (cache durations that were expressed as a mix of milliseconds, seconds, and minutes) I now insist that the framework type `TimeSpan` is used instead for expressing these durations - as that's exactly what it was designed for.