BlockingQueue vs Semaphore

Question

If we want to implement a resource pool such as a database connection pool. Which concurrent collection will you use ? BlockingQueue or Semaphore ?

For BlockingQueue, just like the producer-consumer design pattern, the producer will place all the connections on the queue and the consumer will take next available connection from the queue.

For Semaphore, you specify the semaphore to the pool size, and acquire permit until you reach the pool size and wait for any of them to release the permit and putting the resource back in the pool.

Which one is simpler and easier ? and what are the scenario where we can only use one but not other ?

Answer 1

Many blocking queues are implemented with semaphores anyway, (and maybe a mutex/futex/CS). I use blocking queues for object storage a lot - once you have a blocking queue that works, why bother using anything else for an object pool?

Answer 2

The BlockingQueue is simpler because it will keep track of the Connections/resource as well.

e.g.

public abstract class ResourcePool<Resource> {
    private final BlockingQueue<Resource> free;

    protected ResourcePool(int freeLimit) {
        free = new ArrayBlockingQueue<>(freeLimit);
    }

    public Resource acquire() {
        Resource resource = free.poll();
        return resource == null ? create() : resource;
    }

    public void recycle(Resource resource) {
        if (!free.offer(resource))
            close(resource);
    }

    protected abstract Resource create();

    protected abstract void close(Resource resource);
}

As you can see, the BlockingQueue helps keep track of free resources and ensure there is not too many free resources. It is thread safe without requiring explicit locking.

If you use a Semaphore, you still need to store the resources in a collection (making the semaphore redundant ;)

Answer 3

For an advanced connection pool, i would probably use neither. As @PeterLawrey points out, BlockingQueue makes the most sense for a simple pool where all the resources exist initially. however, if you want to do anything more complex, like create resources on demand, then you'll most likely need additional concurrency constructs. in which case, you'll most likely end up using a simple synchronized block or Lock in the end.