There is a reason indeed.

And the reason is type coercion:

In your coalesce, you enter 0 as a second value. 0 is an integer.

So Spark will coerce this to a decimal type. For not losing any information, it needs 10 digits in front of the comma (max value of a signed integer is 2147483647 -> 10 digits).

So when you put (15,6) you only have 9 digits => spark coerces this to 16,6.

The same for (8,0) => 10,0

And (12,0) remains 12,0.

If you want to avoid this coercion, you can pass 0.00 (decimal) instead of 0 (int).

I totally agree that the type coercion can be confusing btw.