An experiment at building a faster ThreadPoolExecutor

Compared to concurrent.futures.ThreadPoolExecutor:

Better performance and efficiency:

• Much less overhead per task. This means you can usefully parallelize more cases where there are many tasks and/or fast tasks.
• Limit how many tasks are iterated over, even when buffersize is not used, reducing memory usage when lazy task creation is used.

Correctness:

• contextvars context is propagated from the time a task is scheduled, rather than the thread pool creation. See this discussion for details.

More flexibility in API design:

• map(func, *iterables, return_in_order=False) allows for results to be returned when finished, rather than in order. This can speed up execution slightly.