I'm not sure how you can confidently assert that "almost nothing" breaks with Java 9. Maybe we were just unlucky that the very first application (a simple less than 20KLoC JavaFX Gui) we used to test upgrading to Java 9 broke immediately. But actually googling "Java 9 breakage" returns a ton of results so I guess our experience wasn't an outlier.

You are just repeating that it's fundamentally the user's fault because they "should not be doing that". A properly designed tool prevents incorrect usage. If a version of Java allows you to do something and that something proves to be very useful to GET THE JOB DONE, then I'm with Linus - the blame is not with the user.

For example, there was simply no alternative to using Unsafe in the past to get reasonable Java performance for certain task. The fact that it was there allowed me to resist the pressure to move a significant part of our codebase to C++. Of course we "should not be doing that" (using Unsafe) but the alternative was abandoning a large existing code base and hiring a bunch of C++ experts or retraining our existing developers or spending 10 times as much on our server capacity. Using Unsafe was a simple pragmatic decision. You cannot say it was "wrong" without knowing all the factors that went into the decision.

This is why I like Linus' stance in this case.