I think that unions are used too often for clever attempts at "cleaning up" code.

They are originally intended for storing "either" varable A or variable B or variable C in the same memory space.

Greate care - and thought - should be given when instead trying to store "both" variable type A and variable B at the same time, i.e. when trying to use unions for type conversion. The "clever" or "beautiful" code is more fragile, and isn't as simple to read.

In this case, it would be easier to create an enumeration naming the indices of the different array entries, and instead access my_time[MINUTES].

Working with an 8-bit processor will allow some abuse that may change drastically with other processors. A 32-bit processor may add 2 pad bytes to make the data type n*4 bytes large. And if one or more of inidivudal fields are larger than a single byte, then there may be padding between the individual fields in a struct. And for fields larger than a byte, the byte order within the field will matter.

In the end, you normally get the lowest maintainance costs by striving for great transparency. Writing code that does exactly what it looks like it is doing. Yes - sometimes you have to write a couple of extra characters. But the efficiency of the program isn't proportional to the number of characters in the source code.