I usually create a private static inner class and keep it as small as possible at the end of the test class that uses it. Some of these these such as MyCacheListener can be replaced by spy(CacheListener.class) and I always try to convert these where possible.
If it's a simple class that only tests in one package will use (like a custom AssertJ class that nothing outside that package will use) then I make it package-protected and stick it in the same package and srcSet as the tests that will use it. I think something like MonthBasedPartitionResolver is simple enough that it should exist within the same src set and in the same package as the test or a few tests that use it. And if another test in some other src set or in another package needs something similar, then too bad, that other test should create its own simple small class that is similar. This sort of class does NOT belong in a src/main of some module like geode-dunit. Better yet make it a static inner class in every test that uses it. I know some people object to this, but here's why... If I want to modify one test that uses MonthBasedPartitionResolver and this modification involves changing MonthBasedPartitionResolver (which remember it's a simple class) then if each test has its own copy I won't risk breaking other tests. When I'm fixing up, cleaning up and refactoring tests, the biggest nightmares I run into are shared utility classes such that if I try to cleanup one test, I can't -- I have to clean up potentially dozens of tests because they linked some some stupid utility class because "code sharing" is great. MonthBasedPartitionResolver is a small example, there are plenty of other examples that are true monstrosities. Sharing code between tests is NOT great unless you devise a Rule or custom AssertJ Assertion that is truly very reusable across many tests -- and only in this case it belongs in geode-junit or geode-dunit where it should live in a src/main and there should be tests for it under src/test or src/integrationTest. I also highly recommend that you keep shared testing objects like this small in scope with a single high-level responsibility. And please, never create something that simply creates a new improved API so that you don't have to use the Geode API -- this just promotes leaving crappy product APIs in place without changing them because by using some fancy testing API, we don't feel the pain that Users experience with the product APIs. On Fri, Oct 12, 2018 at 3:32 PM, Patrick Rhomberg <prhomb...@apache.org> wrote: > Hello, all! > > There are a number of classes that require some number of "toy" classes > as part of their testing framework, e.g., to be used as custom data > containers. Many of these classes involve testing or use of the `deploy > jar` command, and so are packaged into jars for this purpose. > In an effort to promote good coding habits and, as importantly, consensus > throughout this community and our codebase, I would like to ask which of > these are considered the preferred method to maintain these additional > classes. > I realize a priori that none of these options will be applicable in all > cases, but am curious of how the prioritization of these options are > ordered among you all. > > -- Class definition -- > For definition of the classes consumed, we observe all of the following in > the Geode codebase. > > Option C-1: Toy classes are defined as a proper class in a reasonable > package. > -- suboption (a): The class is defined in the module in which it is > consumed, as a resource. [1] > -- suboption (b): The class is defined in the module in which it is > consumed, as a neighboring file. [2] > -- suboption (c): The class is defined in geode-dunit or geode-junit. [3] > > Option C-2: Toy classes are defined as inner classes of the test class. > [4] > > Option C-3: Sufficiently small toy classes are defined as raw String > fields in the test class that consumed it and is compiled by the test JVM > via the JavaCompiler class. [5] > > -- Resource exposure -- > For those tests that require classes placed in Jars for use in the `deploy > jar` command, we observe the following in the Geode codebase: > > Option J-1: Jars are a resource and are built or downloaded as necessary > during the build steps, before test execution. [6] > > Option J-2: Sufficiently small classes that are defined as raw String > fields in the test class that consumes it (C-3 above) are packaged into > resources via the JarBuilder class. [7] > > > I look forward to hearing your opinions. > > Imagination is Change. > ~Patrick Rhomberg > > Examples: > [1] > geode-core_test/resources:org.apache.geode.management.internal.deployment. > ConcreteExtendsAbstractExtendsFunctionAdapter.java > is found by the FunctionScannerTest to be a user-defined function. > [2] > geode-core_distributedTest:org.apache.geode.internal. > cache.SerializableMonth > defines a simple DataSerializable containing an int for the month, to be > consumed by *.MonthBasedPartitionResolver > in PRCustomPartitioningDistributedTest. > [3] geode-junit_main:org.apache.geode.cache.query.transaction.Person > defines a simple DataSerilazable container class for a String name and int > age. > [4] > geode-core_distributedTest:java.org.apache.geode.internal.statistics. > StatisticsDistributedTest$PubSubStats > is consumed by its parent class > [5] > geode-assembly_acceptanceTest:org.apache.geode.management. > internal.cli.commands.PutCommandWithJsonTest > defines a Customer class in a raw string in the test's @Before method. > [6] See our own consumption of geode-dependencies.jar et al via the > StartMemberUtils.java, or our historic consumption of tools.jar > [7] See again [5] or other uses of the JarBuilder class. >
