On 3/1/19 3:49 PM, Matt Pulver wrote: > In many applications, I would much rather see calculations carried out > via IEEE 754 all the way to the end, with nans and infs, which > provides much more useful diagnostic information than an exception that > doesn't return any rows at all. As Andres Freund pointed out, it is also > more expensive to do the intermediate checks. Just let IEEE 754 do its > thing! (More directed at the SQL standard than to PostgreSQL.)
I wanted to try this out a little before assuming it would work, and there seems to be no trouble creating a trivial domain over float8 (say, CREATE DOMAIN ieeedouble AS float8), and then creating operators whose operand types are the domain type. So it seems an extension could easily do that, and supply happily inf-returning and NaN-returning versions of the operators and functions, and those will be used whenever operands have the domain type. It might even be useful and relatively elegant, while leaving the SQL-specified base types to have the SQL-specified behavior. -Chap