How many Futures is too many in Java?

You're conflating a Future, which represents the possible result of an asynchronous operation with a Thread which represents the ability to perform processing on a Callable (in the case of an Executor at least).

There's nothing to stop you calling submit on a thread pool millions of times and get a huge list of Future objects for you to wait on. You don't even need to wait for them to finish if the application will continue running and you have no need to process the result.

But.

If you create all these jobs, they are going to require memory to hold their state. If that memory is somehow part of the input to the job, or the result of executing the job, then you will commit heap space to all these tasks. You can't do this forever. Essentially, you need to think of some sort of throttling, if you're going to pull huge amount of work into a process to run in the background.

To my understanding, ExecutorServices with a pool of heavyweight threads, handles Futures like they are green thread

That's not correct. If we ignore the bells and whistles, an ExecutorService consists of a collection of worker threads, and a blocking queue of tasks. Each task in the queue is a wrapper containing one of your tasks, and a Future.

Each worker thread loops forever,

• Picks a task from the queue,
• Calls your Runnable or Callable object's run() or call(...) method,
• Completes the Future with the value returned by your method or, with an exception that was thrown by your method.
• Goes back to wait for another task.

The only threads are the "heavy weight" worker threads. Once one of the worker threads starts to work on a task, it won't do anything else until the task is complete. Tasks that haven't yet been started are just objects in a queue, and the Executor forgets about each task and Future object as soon as the Future is completed. Those won't continue exist after your own code has discarded the references to them.

Should I submit millions of Futures to ExecutorService?

you can, but you should evaluate possible time overhead. The overhead of handling separate Future object is small, but greater than zero. So the less number of tasks, the better. On the other hand, when the number of tasks becomes less then the number of processors (that is, the number of processor cores with respect to hyperthreading), then the level of parallelism is reduced and the overall execution time increases.

Let you have 1 million of 10-ms tasks, and your computer have 8 cores. Then the overall execution time of 1250 sec is increased by (10*8/2) = 40 ms because of decreasing parallelism at the end, and plus 125 ms for task switch (I evaluate it as little as 1 useq for each task switch). if your have 100000 100-ms tasks, then the execution time is still expected to be 1250 sec, plus 400 ms for tail and 12.5 ms for switch. Either way, the time overhead is neglidgible, but it can increase if your tasks are significantly shorter or longer than 10...100 interval.