limitedParallelism does not propagate the full CoroutineContext to the underlying dispatcher

[DmitryTsyvtsyn]

No description provided.

[dkhalanskyjb]

Issue: #4305

This approach doesn't work, and the issue is far from being this easy to solve. See this code:

val limited = object : CoroutineDispatcher() {
    override fun dispatch(context: CoroutineContext, block: Runnable) {
        Dispatchers.Default.dispatch(EmptyCoroutineContext, Runnable {
            println("${Thread.currentThread()}: Ensuring the context $context")
            block.run()
            println("${Thread.currentThread()}: Finished executing in the context $context")
        })
    }
}.limitedParallelism(2)

limited.dispatch(CoroutineName("Hi!")) {
    Thread.sleep(100)
    println("${Thread.currentThread()}: Hello!")
}
limited.dispatch(CoroutineName("Servus!")) {
    Thread.sleep(100)
    println("${Thread.currentThread()}: Hallo!")
}
limited.dispatch(CoroutineName("Hola!")) {
    Thread.sleep(100)
    println("${Thread.currentThread()}: ¡Hola!")
}
limited.dispatch(CoroutineName("Привет!")) {
    Thread.sleep(100)
    println("${Thread.currentThread()}: Здравствуйте!")
}

Output with the proposed change:

Thread[DefaultDispatcher-worker-2,5,main]: Ensuring the context CoroutineName(Servus!)
Thread[DefaultDispatcher-worker-1,5,main]: Ensuring the context CoroutineName(Hi!)
Thread[DefaultDispatcher-worker-2,5,main]: Hallo!
Thread[DefaultDispatcher-worker-1,5,main]: Hello!
Thread[DefaultDispatcher-worker-2,5,main]: ¡Hola!
Thread[DefaultDispatcher-worker-1,5,main]: Здравствуйте!
Thread[DefaultDispatcher-worker-2,5,main]: Finished executing in the context CoroutineName(Servus!)
Thread[DefaultDispatcher-worker-1,5,main]: Finished executing in the context CoroutineName(Hi!)

We see that "Hola" and "Здравствуйте" run in the wrong context anyway. What's needed here if we decide to fix the issue (which may well not be worth the effort) is a more complex setup that checks which workers run in which contexts and ensures that each task runs in the context appropriate for it, possibly shutting workers down and creating new ones to allow that.