On 01/27/2016 11:16 AM, Aurelien Pupier wrote:
About using fragments, the only drawback I'm aware of is that one cannot add a dependency to a fragment from their plugin. So it's not possible for other tests to reuse some logic that is inside a fragment. In the spirit of "treating tests as 1st class citizen", I believe it's generally better to allow composition of test bundles and to make them regular bundles.

If reusable, the logic can be moved to a dedicated utility test plugin.
Yes, but it requires an effort in moving the code to the right plugin or turning fragment into a bundle everytime one identifies something to reuse; whereas using bundles only doesn't require any effort from the "producer" side.

On the other hand, some integration tests should be able to detect that.
That would require writing some additional integration tests, whereas the current approach to run "unit" tests inside a workbench provide that verification without additional thing to write.

The QE people could tell you their opinion on what's more important between fast tests, and longer tests that are closer from production.
I guess it all depends, as always: we could imagine the unit test running in plain Java to verify only their logic, but it shouldn't become a replacement to some integration test to verify that the logic also works in the right context.
Not planned to have them as a replacement. I already created the structure to write also higher integration tests.
If, during automated tests/build, you run both unit tests and then start a workbench to run integration tests, then you pay the price of the workbench startup anyway. So why not running those unit tests in the workbench?

When running locally, I may want to run only unit tests first.  Without launching the OSGi platform it will be faster. It will also allow to use some tools such as Infinitest to have continuous feedback while developing.
I don't know much of Infinitest, but I believe it doesn't rely on how Maven runs tests. Eclipse already provide the ability to run a test class in a bundle in plain Java without starting the workbench; it's "Run As > JUnit Test" instead of "Run As > JUnit Plugin Test". I guess Infinitest can rely on that, can't it?

Just to be clear, I'm not saying it's wrong to change that; I'm more wonderint how much profitable is it, what changing test structure would provide better than the current one does. The Infinitest story seems the only one "worth it" IMO, and it doesn't seem to be correlated to tycho-surefire vs maven-surefire.
Mickael Istria
Eclipse developer at JBoss, by Red Hat
My blog - My Tweets