Codegen compilation error for Float scalars with default values in Input objects

[joaquim-verges]

Summary

We recently upgraded from Apollo Android 2.4.1 to 2.5.6, and encountered a compilation error in the autogenerated code.

In short, Input objects in our json schema with inputs of type Float and default values that omit the decimal part (eg. "defaultValue": "1") causes the Java codegen to generate a default value as an Integer instead of Double which causes a compilation error:

private Input<Double> myFloatField = Input.fromNullable(1);
                                                        ^
equality constraints: Double
lower bounds: Integer

Version
2.5.6

Description

I was able to narrow down the issue to a specific codegen corner case:

• Only happens in Java codegen (issue does not happen with Kotlin models)
• Only happens with json schema format (issue does not happen with sdl schema format)

When a INPUT_OBJECT type has a Float input field with a default value of 1 or 0 (omitting the decimal part) like so:

 {
  "kind": "INPUT_OBJECT",
  "name": "ReviewRefInput",
  "description": "Circle ref to review input",
  "fields": null,
  "inputFields": [
    {
      "name": "myFloatField",
      "description": "Testing float input codegen",
      "type": {
        "kind": "SCALAR",
        "name": "Float",
        "ofType": null
      },
      "defaultValue": "1"
   }
  ]
}

The corresponding output in the generated java file is:

private Input<Double> myFloatField = Input.fromNullable(1);

Which is a compilation error, as 1 is considered an Integer, and therefore that assignment fails with:

private Input<Double> myFloatField = Input.fromNullable(1);
                                                        ^
equality constraints: Double
lower bounds: Integer

The equivalent Kotlin generated code is correct however:

val myFloatField: Input<Double> = Input.optional(1.0)

The Kotlin codegen correctly converts 1 to 1.0 since the scalar is type Float, but the Java codegen does not.

This also happens for mandatory fields:

java: private double mandatoryFloat = 1;
kotlin: val mandatoryFloat: Double = 1.0

Interestingly, the java codegen does the correct transformation when using schemas in sdl format, so the issue only happens when parsing from json.

I was able to reproduce the issue easily with the apollo-compiler unit tests by modifying the colorInput entry in src/test/validation/schema.json, changing the green input field to have a default value of 0 instead of 0.0.

The generated file ColorInput.java shows the error:

private Input<Double> green = Input.fromNullable(0);

while the equivalent kotlin file is correct:

val green: Input<Double> = Input.optional(0.0)

Hope this is clear enough, let me know if you need any more information, and thanks for this amazing library!

[joaquim-verges]

A little more information:

We first checked if this was a schema definition issue, but the spec says it is a valid case, and we confirmed it here.

I looked a bit deeper at the code, and believe the issue was introduced in #2917 specifically in GraphQLParser.ValueContext.toKotlin() that was added to IRBuilder.kt l. 152:

is GraphQLParser.NumberValueContext -> if (
          NUMBER().text.contains('.') ||
          NUMBER().text.contains("e")) {
        NUMBER().text.toDouble()
      } else {
        NUMBER().text.toInt()
      }

Which shows exactly what the problem is.

I was able to fix it by passing the IntrospectionSchema.TypeRef to that method and check the scalar type to return the correct type.

I'm going to push a PR for it :)

[joaquim-verges]

Pushed a fix in #3109 - let me know if I got it right :)