I've been looking at key-value stores recently, and I wonder what people think about collating numbers of heterogeneous types. FoundationDB tuples have type co

I've been looking at key-value stores recently, and I wonder what people think about collating numbers of heterogeneous types. FoundationDB tuples have type codes that segregate values of different types, so that the strings "1" and "2" sort before the integers 1 and 2, which sort before the single-precision floats 1.0f and 2.0f, which sort before the double-precision floats 1.0 and 2.0. The database that I test supports double-precision IEEE floats, and a proprietary decimal float with a signed 64-bit significand and signed 8-bit exponent. When converted to string for use as keys, these collate as expected. The price of this is that you don't get shortest representations of the sort sought by this paper and others. Otherwise, a binary and decimal float that compare unequal could convert to the same string. I guess it's kind of unusual to use floats as keys, and more unusual still to use both binary and decimal floats, but I wonder if there is another strategy for collating them.