Page History

Distributed Test Recommendations

The goals of the Green Team

• Improve green consistency of distributed tests
• Mass test runs determined which tests we worked on
• Prioritized any tests that failed more than once in our mass test runs
• Fixed about 35 issues (working on more)
• Spent about a quarter of our time fixing product bugs
• Anecdotal improvements for Geode developers
• “CIO board is cleaner”
• “Seeing more Green pull requests”

...

If I had a nickel for every thread sleep related failure...

What was helpful vs unhelpful

...

• All test code should be in the test class
• Test should directly use Geode User APIs without helper classes
• Most Geode developers already know the Geode APIs or should learn them
• See how Geode is configured and what’s happening directly in the test code
• Easier to debug especially if product bugs are suspected
• Developers understand usability issues in the Geode User APIs
• Every layer between the test and what you are testing obscures the test and make debugging more difficult for you and everyone else
• Avoid generalizing classes such as CacheListeners for many tests
  • Use specialized inner classes that are specific to the tests in the test class
  • Flexibility and generalization are sources of complexity
  • Complexity is a source of bugs and difficult debugging
  • Complexity and generalization is a wall that prevents people from making educated changes
  • All code including configuration should be in the test class
  • Write your code for the next person to look at it.
• Helper classes cause problems
  • Obscure what’s happening
  • Proliferate bad anti-patterns
  • Combine test classes that should be separate

Challenges when handling things asynchronously...

Use ExecutorServiceRule for multithreading

• Use ExecutorServiceRule or DistributedExcutorServiceRule
• Provides debugging support for hangs
• Cleans up threads on tear down
• Test task code should be interruptible
• Use Future or CompletableFuture for submitted runnable/callable
• Always invoke get() on any Future or CompletableFuture

...

• Avoid catching exception and setting some test state to check later
• Use ErrorCollector in unit and integration tests
• Use SharedErrorCollector in distributed tests

AsyncInvocation usage

...

• DistributedRule -- simply launches DUnit and greps for suspect strings after each test method
• DistributedExecutorServiceRule -- provides an ExecutorService for all VMs
• DistributedRestoreSystemProperties -- restores system properties in all VMs
• SharedCountersRule -- shares counters across all VMs
• SharedErrorCollector -- shares one JUnit ErrorCollector across all VMs

...

• Use clean code with meaningful names
• Describe the point of the test in a comment
• Use Awaitility and CountDownLatches instead of sleeping/pausing
• Use AssertJ instead of JUnit Assert
• Use Geode defaults or configure using Geode APIs directly within the test
• Use parameterization instead of test class inheritance
• Avoid sharing callback or domain classes across multiple tests
• Use ErrorCollector for assertions and exception handling in callbacks instead of rethrowing or setting state to check later
• Use factory based injection instead of mutable/static product test hooks
• Use Geode user APIs instead of adapter/helper/junit-rule APIs
• Avoid unnecessary use of invokeAsync or multithreading
• Use AsyncInvocation or Future to wait when using invokeAsync
  • Use AsyncInvocation.await() for void types
  • Use AsyncInvocation.get() for non-void types
• Use only non-deprecated Geode user APIs in tests (unless the test is specifically for deprecated APIs)
• Use custom logging when debugging hard-to-reproduce problems
• Know if the call you are making is asynchronous and wait for it, you would be surprised at how many are asynchronous

...