Sqrt

[NicoLaval]

Sqrt throws errors for calls with null / NaN / negative values, BdI tests assume null to be the correct result.

[NicoLaval]

@noahboerger

• NaN : sqrt only admits number as parameter, seems to be the good behaviour to throw an exception
• negative: Trevas bug , has to return value
• null:
  • scalars : sqrt(null) null has no inferred type, you have to write sqrt(cast(null, number))
  • components / datasets:
    Considering ds_1 := ds[calc sq := sqrt(a)]; and ds:

id	a
IDENTIFIER	MEASURE
INTEGER	INTEGER
1	2
2	-2
3	null

Trevas ds_1 result will be:

id	a	sq
IDENTIFIER	MEASURE	MEASURE
INTEGER	INTEGER	INTEGER
1	2	4
2	-2	4
3	null	null

Thanks to dataset metadata, null is recognized as integer.

[NicoLaval]

TODO: fix behaviour for negative inputs

[noahboerger]

Thank you for the explanations and plan to fix it for negative inputs.

[hadrienk]

Just a comment here, but I think I've seen somewhere in the spec that type inference should be supported. It's a very small paragraph and there's not much details. Technically I think this isn't too hard to implement and has great value for the users. But this need to be reworked in the spec.

[NicoLaval]

Not sure about that.
What will be rewrited in 2.1 is implicit casting between known types.

I think your right, when there is no doubt about the null type, we can apply an implicit cast, avoiding users to write lot of cast.

I opened an issue in the TF repo