Skip to content

Latest commit

 

History

History
171 lines (141 loc) · 11.9 KB

README.md

File metadata and controls

171 lines (141 loc) · 11.9 KB

Fixtures CI

Fixtures is a library that helps us to instantiate data classes for our tests easily. We were inspired by this blog post, and decided to use KSP to automate the generation of the described functions.

Installation

Before using this library, we must set up KSP in our project. We can follow the instructions here. Then we must include the following dependency:

implementation("io.github.bluegroundltd:fixtures-annotations:1.0.7")
implementation("io.github.bluegroundltd:fixtures:1.0.7")

How to

All we have to do is to add the Fixture annotation in a data class. For example, if we have the following data class:

@Fixture
data class Foo(
     val stringValue: String,
     val intValue: Int,
)

Then the following helper function will be generated:

fun createFoo(
     stringValue: String = "stringValue",
     intValue: Int = 0
) : Foo = Foo(
     stringValue = stringValue,
     intValue = intValue
)

The naming convention for the generated function is the class' name with the create prefix. So, from the Foo class, the createFoo function will be generated. This function will be placed in a file with the FooFixture.kt name, and the file will be placed in the Foo's package under the build/generated/ksp/kotlin/ path.

Randomize your data

The generated functions have default values for their parameters. The values are standard based on the parameter's type. If we want to assign random values then we can use the next KSP option:

ksp.arg("fixtures.randomize", "true")

After applying the previous option, every time we generate the functions new default values will be assigned. Things to note here are:

  • The default behavior is to not randomize the data.
  • The randomization happens in every generated function in the Gradle module. In the future, we may consider randomizing data per fixture.

Supported field types

The supported field types are:

Supported type default value randomized default value
String A string whose value will be equal to the name of the field An at most 10 characters long string that contains random alphanumeric character
Char 'A' A random alphanumeric character
Boolean false Random.nextBoolean()
Int 0 Random.nextInt(20)
Long 0L Random.nextLong( 20)
Float 0f Random.nextFloat()
Double 0.0 Random.nextDouble(20.0)
Date Date(0) Date()
TimeZone TimeZone.getTimeZone("UTC") TimeZone.getTimeZone(TimeZone.getAvailableIDs().random())
UUID UUID.fromString("00000000-0000-0000-0000-000000000000") UUID.randomUUID()
LocalDate LocalDate.of(1989, 1, 23) LocalDate.now()
LocalTime LocalTime.of(0, 0) LocalTime.now()
LocalDateTime LocalDateTime.of(1989, 1, 23, 0, 0) LocalDateTime.now()
ZonedDateTime ZonedDateTime.of(1989,1,23,0,0,0,0, ZoneId.of("UTC")) ZonedDateTime.now()
Instant Instant.EPOCH Instant.now()
OffsetTime OffsetTime.MIN OffsetTime.now()
OffsetDateTime OffsetDateTime.MIN OffsetDateTime.now()
ZoneId ZoneId.of("UTC") ZoneId.of(ZoneId.getAvailableZoneIds().random())
BigDecimal BigDecimal.ZERO BigDecimal.valueOf(Random.nextDouble(20.0))
BigInteger BigInteger.ZERO BigInteger.valueOf(Random.nextInt(20))
Array emptyArray() emptyArray()
List emptyList() emptyList()
Set emptySet() emptySet()
Map emptyMap() emptyMap()
Enums The first enum entry Randomly selected enum entry
Sealed classes (their data subclasses should be annotated with @Fixture) The first sealed class' entry Randomly selected sealed class' entry
Other data classes annotated with @Fixture Invocation to other @Fixture's generate function Invocation to other @Fixture's generate function

Handle not supported field types

There will be some rare cases where we have a data class that contains another class which does not belong in the supported field types. Additionally, this not supported class may be part of another library, so we could not use the Fixture annotation on it. Let's assume that we have the following class:

data class Foo(
     val doubleValue: Double,
     val barValue: Bar
)

and that Bar class is part of another library. To help the processor generate a Bar class we can use the FixuteAdapter annotation. This annotation can be applied to a top-level function, like that:

@FixtureAdapter
fun barFixtureProvider(): Bar = Bar()

Then the processor will use the annotated function to generate the helper function.

fun createFoo(
     doubleValue: Double = 0.0,
     barValue: Bar = barFixtureProvider()
) : Foo = Foo(
     doubleValue = doubleValue,
     barValue = barValue
)

Unfortunately, the functions that are annotated with the FixtureAdapter annotation must be placed into the main source sets, till the source set issue is resolved.

Source sets Issue

For now, we can not run KSP on the main source sets and generate classes in the test source sets. This seems that will not be the case after fixing this issue. Until then, we can avoid generating test functions in our release code by setting the following KSP option:

// Instead of setting statically the value to false you 
// should create a function to calculate this value.
ksp.arg("fixtures.run", "false") 

The default value of this option is true. We need to explicitly set it to false when we are about to run our tests. This can be easily done with a function in our Gradle file.

Multi-module support

Fixtures in different modules

If our data class contains a Fixture field, and the declaration of the field's class belongs to another module, then our processor can not recognize that it is a Fixture field. The reason is that KSP seems to not provide the annotations of a class defined in another module. To help our processor we added need to use the ModularizedFixture annotation.

For example, if we have defined in one module the following class

data class Foo(val barValue: Bar)

and in another module this class

@Fixture
data class Bar(val stringValue: String)

then to help our processor to generate the fixture, we need to annotate the field with the ModularizedFixture annotation

@Fixture
data class Foo(@ModularizedFixture val barValue: Bar)

Sealed classes in different modules

If our data class contains a sealed class field, and the declaration of the sealed class belongs to another module, then our processor can not recognize that this field is a sealed class. This means that it can not treat it accordingly. The reason behind that is that the generated bytecode does not contain any information about being a sealed class (due to java interoperability). To overcome this issue we can use FixtureAdapter as described in this section.