Yeah, poor performance and memory usage. In the Java world, thread locals are java.lang.ThreadLocal, basically a hashmap from thread to variable.

Using a ThreadLocal is (usually) a code smell. It's optimized / supposed to be used with just a few threads, not millions.

For what it worth, there actually is a separate JEP (I believe this: https://openjdk.org/jeps/429 ) for a new, scope-bases solution that promises much better performance.

But the lock would only need to be per platform thread?