set the channel topic: https://github.com/MeihanLi/pinot/pull/1

set the channel description: https://github.com/MeihanLi/pinot/pull/1

thanks

folks, created a channel to discuss https://github.com/MeihanLi/pinot/pull/1

we see more and more users run into this issue

and it's a concerning one, as it could return incorrect result silently

any suggestions on having a stop-gap sooner?

I agree that this is concerning, although I am not quite sure if there is a good enough stop-gap fix that is available. Seems to me that any stop-gap fix would still suffer from some of the type issues that we see now. So, at the very least it would be good to know what our final solution would be to fix this issue before we start adding stop-gap fixes.

makes sense. we can discuss the ideal state, and then see if there's opportunity to have some easy solution to approximate

add @User @User too, since i brought this up a while ago to them

my thoughts is to have a typing hierarchy from high precision to low precision. so any high precision will short-circuit and map to its corresponding implementation

Ideally we should pick the same way as how java performs the type matching (high to low precision, and implicit type conversion)

good point

we need to implement this implicit cast rule https://docs.oracle.com/cd/B19306_01/server.102/b14200/sql_elements002.htm#g195937

we would also need to detect if implicit conversion is lossy. For example, while java allows implicit conversion of

long

to

double

, that conversion is lossy.

public static void main(String[] args) {
  long l = 8223372036854775802l;
  double d = l;
  long f = (long) d;

  System.out.println("Original Long Value      : " + l);
  System.out.println("Intermediate Double Value: " + d);
  System.out.println("Final Long Value         : " + f);

  System.out.println( l == f ? "Not a lossy conversion" : "Lossy conversion");
}

If we convert a

long

value to

double

and then convert the resulting

double

value back to

long

, then the final

long

value may not be same as initial

long

value due to lossy conversion of

long

to

double

. Currently the output type of all functions in

ArithmeticFunctions.java

is set to

double

.

interesting

also add @User to this channel, as there was a similar conv on a PR review

Adding @User as well, since he had a few thoughts regarding scalar functions using templates.