How to make ThreadPoolExecutor's submit() method block if it is saturated?

I want to create a ThreadPoolExecutor such that when it has reached its maximum size and the queue is full, the submit() method blocks when trying to add new tasks. Do I need to implement a custom RejectedExecutionHandler for that or is there an existing way to do this using a standard Java library?

One of the possible solutions I've just found:

public class BoundedExecutor {
    private final Executor exec;
    private final Semaphore semaphore;

    public BoundedExecutor(Executor exec, int bound) {
        this.exec = exec;
        this.semaphore = new Semaphore(bound);
    }

    public void submitTask(final Runnable command)
            throws InterruptedException, RejectedExecutionException {
        semaphore.acquire();
        try {
            exec.execute(new Runnable() {
                public void run() {
                    try {
                        command.run();
                    } finally {
                        semaphore.release();
                    }
                }
            });
        } catch (RejectedExecutionException e) {
            semaphore.release();
            throw e;
        }
    }
}

Are there any other solutions? I'd prefer something based on RejectedExecutionHandler since it seems like a standard way to handle such situations.

You can use ThreadPoolExecutor and a blockingQueue:

public class ImageManager {
    BlockingQueue<Runnable> blockingQueue = new ArrayBlockingQueue<Runnable>(blockQueueSize);
    RejectedExecutionHandler rejectedExecutionHandler = new ThreadPoolExecutor.CallerRunsPolicy();
    private ExecutorService executorService =  new ThreadPoolExecutor(numOfThread, numOfThread, 
        0L, TimeUnit.MILLISECONDS, blockingQueue, rejectedExecutionHandler);
	
    private int downloadThumbnail(String fileListPath){
        executorService.submit(new yourRunnable());
    }
}

You should use the CallerRunsPolicy, which executes the rejected task in the calling thread. This way, it can't submit any new tasks to the executor until that task is done, at which point there will be some free pool threads or the process will repeat.

http://java.sun.com/j2se/1.5.0/docs/api/java/util/concurrent/ThreadPoolExecutor.CallerRunsPolicy.html

From the docs:

> Rejected tasks

> New tasks submitted in method execute(java.lang.Runnable) will be > rejected when the Executor has been > shut down, and also when the Executor > uses finite bounds for both maximum > threads and work queue capacity, and > is saturated. In either case, the > execute method invokes the > RejectedExecutionHandler.rejectedExecution(java.lang.Runnable, > java.util.concurrent.ThreadPoolExecutor) > method of its > RejectedExecutionHandler. Four > predefined handler policies are > provided: > > 1. In the default ThreadPoolExecutor.AbortPolicy, the > handler throws a runtime > RejectedExecutionException upon > rejection. > 2. In ThreadPoolExecutor.CallerRunsPolicy, > the thread that invokes execute itself > runs the task. This provides a simple > feedback control mechanism that will > slow down the rate that new tasks are > submitted. > 3. In ThreadPoolExecutor.DiscardPolicy, a > task that cannot be executed is simply > dropped. > 4. In ThreadPoolExecutor.DiscardOldestPolicy, > if the executor is not shut down, the > task at the head of the work queue is > dropped, and then execution is retried > (which can fail again, causing this to > be repeated.)

Also, make sure to use a bounded queue, such as ArrayBlockingQueue, when calling the ThreadPoolExecutor constructor. Otherwise, nothing will get rejected.

Edit: in response to your comment, set the size of the ArrayBlockingQueue to be equal to the max size of the thread pool and use the AbortPolicy.

Edit 2: Ok, I see what you're getting at. What about this: override the beforeExecute() method to check that getActiveCount() doesn't exceed getMaximumPoolSize(), and if it does, sleep and try again?

I know, it is a hack, but in my opinion most clean hack between those offered here ;-)

Because ThreadPoolExecutor uses blocking queue "offer" instead of "put", lets override behaviour of "offer" of the blocking queue:

class BlockingQueueHack<T> extends ArrayBlockingQueue<T> {

	BlockingQueueHack(int size) {
		super(size);
	}

	public boolean offer(T task) {
		try {
			this.put(task);
		} catch (InterruptedException e) {
			throw new RuntimeException(e);
		}
		return true;
	}
}

ThreadPoolExecutor tp = new ThreadPoolExecutor(1, 2, 1, TimeUnit.MINUTES, new BlockingQueueHack(5));

I tested it and it seems to work. Implementing some timeout policy is left as a reader's exercise.

Hibernate has a BlockPolicy that is simple and may do what you want:

See: Executors.java

/**
 * A handler for rejected tasks that will have the caller block until
 * space is available.
 */
public static class BlockPolicy implements RejectedExecutionHandler {

	/**
     * Creates a <tt>BlockPolicy</tt>.
     */
    public BlockPolicy() { }

    /**
     * Puts the Runnable to the blocking queue, effectively blocking
     * the delegating thread until space is available.
     * @param r the runnable task requested to be executed
     * @param e the executor attempting to execute this task
     */
    public void rejectedExecution(Runnable r, ThreadPoolExecutor e) {
    	try {
			e.getQueue().put( r );
		}
		catch (InterruptedException e1) {
			log.error( "Work discarded, thread was interrupted while waiting for space to schedule: {}", r );
		}
    }
}

The BoundedExecutor answer quoted above from Java Concurrency in Practice only works correctly if you use an unbounded queue for the Executor, or the semaphore bound is no greater than the queue size. The semaphore is state shared between the submitting thread and the threads in the pool, making it possible to saturate the executor even if queue size < bound <= (queue size + pool size).

Using CallerRunsPolicy is only valid if your tasks don't run forever, in which case your submitting thread will remain in rejectedExecution forever, and a bad idea if your tasks take a long time to run, because the submitting thread can't submit any new tasks or do anything else if it's running a task itself.

If that's not acceptable then I suggest checking the size of the executor's bounded queue before submitting a task. If the queue is full, then wait a short time before trying to submit again. The throughput will suffer, but I suggest it's a simpler solution than many of the other proposed solutions and you're guaranteed no tasks will get rejected.

The following class wraps around a ThreadPoolExecutor and uses a Semaphore to block then the work queue is full:

public final class BlockingExecutor { 
	
	private final Executor executor;
	private final Semaphore semaphore;
	
	public BlockingExecutor(int queueSize, int corePoolSize, int maxPoolSize, int keepAliveTime, TimeUnit unit, ThreadFactory factory) {
		BlockingQueue<Runnable> queue = new LinkedBlockingQueue<Runnable>();
		this.executor = new ThreadPoolExecutor(corePoolSize, maxPoolSize, keepAliveTime, unit, queue, factory);
		this.semaphore = new Semaphore(queueSize + maxPoolSize);
	}
	
	private void execImpl (final Runnable command) throws InterruptedException {
		semaphore.acquire();
		try {
			executor.execute(new Runnable() {
				@Override
				public void run() {
					try {
						command.run();
					} finally {
						semaphore.release();
					}
				}
			});
		} catch (RejectedExecutionException e) {
			// will never be thrown with an unbounded buffer (LinkedBlockingQueue)
			semaphore.release();
			throw e;
		}
	}
	
	public void execute (Runnable command) throws InterruptedException {
		execImpl(command);
	}
}

This wrapper class is based on a solution given in the book Java Concurrency in Practice by Brian Goetz. The solution in the book only takes two constructor parameters: an Executor and a bound used for the semaphore. This is shown in the answer given by Fixpoint. There is a problem with that approach: it can get in a state where the pool threads are busy, the queue is full, but the semaphore has just released a permit. (semaphore.release() in the finally block). In this state, a new task can grab the just released permit, but is rejected because the task queue is full. Of course this is not something you want; you want to block in this case.

To solve this, we must use an unbounded queue, as JCiP clearly mentions. The semaphore acts as a guard, giving the effect of a virtual queue size. This has the side effect that it is possible that the unit can contain maxPoolSize + virtualQueueSize + maxPoolSize tasks. Why is that? Because of the semaphore.release() in the finally block. If all pool threads call this statement at the same time, then maxPoolSize permits are released, allowing the same number of tasks to enter the unit. If we were using a bounded queue, it would still be full, resulting in a rejected task. Now, because we know that this only occurs when a pool thread is almost done, this is not a problem. We know that the pool thread will not block, so a task will soon be taken from the queue.

You are able to use a bounded queue though. Just make sure that its size equals virtualQueueSize + maxPoolSize. Greater sizes are useless, the semaphore will prevent to let more items in. Smaller sizes will result in rejected tasks. The chance of tasks getting rejected increases as the size decreases. For example, say you want a bounded executor with maxPoolSize=2 and virtualQueueSize=5. Then take a semaphore with 5+2=7 permits and an actual queue size of 5+2=7. The real number of tasks that can be in the unit is then 2+5+2=9. When the executor is full (5 tasks in queue, 2 in thread pool, so 0 permits available) and ALL pool threads release their permits, then exactly 2 permits can be taken by tasks coming in.

Now the solution from JCiP is somewhat cumbersome to use as it doesn't enforce all these constraints (unbounded queue, or bounded with those math restrictions, etc.). I think that this only serves as a good example to demonstrate how you can build new thread safe classes based on the parts that are already available, but not as a full-grown, reusable class. I don't think that the latter was the author's intention.

you can use a custom RejectedExecutionHandler like this

ThreadPoolExecutor tp= new ThreadPoolExecutor(core_size, // core size
        		max_handlers, // max size 
        		timeout_in_seconds, // idle timeout 
				TimeUnit.SECONDS, queue, new RejectedExecutionHandler() {
					public void rejectedExecution(Runnable r, ThreadPoolExecutor executor) {
						// This will block if the queue is full
						try {
							executor.getQueue().put(r);
						} catch (InterruptedException e) {
							System.err.println(e.getMessage());
						}
						
					}
				});

I don't always like the CallerRunsPolicy, especially since it allows the rejected task to 'skip the queue' and get executed before tasks that were submitted earlier. Moreover, executing the task on the calling thread might take much longer than waiting for the first slot to become available.

I solved this problem using a custom RejectedExecutionHandler, which simply blocks the calling thread for a little while and then tries to submit the task again:

public class BlockWhenQueueFull implements RejectedExecutionHandler {

    public void rejectedExecution(Runnable r, ThreadPoolExecutor executor) {

        // The pool is full. Wait, then try again.
        try {
            long waitMs = 250;
            Thread.sleep(waitMs);
        } catch (InterruptedException interruptedException) {}

        executor.execute(r);
    }
}

This class can just be used in the thread-pool executor as a RejectedExecutinHandler like any other, for example:

executorPool = new ThreadPoolExecutor(1, 1, 10,
                                      TimeUnit.SECONDS, new SynchronousQueue<Runnable>(),
                                      new BlockWhenQueueFull());

The only downside I see is that the calling thread might get locked slightly longer than strictly necessary (up to 250ms). Moreover, since this executor is effectively being called recursively, very long waits for a thread to become available (hours) might result in a stack overflow.

Nevertheless, I personally like this method. It's compact, easy to understand, and works well.

Create your own blocking queue to be used by the Executor, with the blocking behavior you are looking for, while always returning available remaining capacity (ensuring the executor will not try to create more threads than its core pool, or trigger the rejection handler).

I believe this will get you the blocking behavior you are looking for. A rejection handler will never fit the bill, since that indicates the executor can not perform the task. What I could envision there is that you get some form of 'busy waiting' in the handler. That is not what you want, you want a queue for the executor that blocks the caller...

To avoid issues with @FixPoint solution. One could use ListeningExecutorService and release the semaphore onSuccess and onFailure inside FutureCallback.

Recently I found this question having the same problem. The OP does not say so explicitly, but we do not want to use the RejectedExecutionHandler which executes a task on the submitter's thread, because this will under-utilize the worker threads if this task is a long running one.

Reading all the answers and comments, in particular the flawed solution with the semaphore or using afterExecute I had a closer look at the code of the ThreadPoolExecutor to see if there is some way out. I was amazed to see that there are more than 2000 lines of (commented) code, some of which make me feel dizzy. Given the rather simple requirement I actually have --- one producer, several consumers, let the producer block when no consumers can take work --- I decided to roll my own solution. It is not an ExecutorService but just an Executor. And it does not adapt the number of threads to the work load, but holds a fixed number of threads only, which also fits my requirements. Here is the code. Feel free to rant about it :-)

package x;

import java.util.concurrent.BlockingQueue;
import java.util.concurrent.Executor;
import java.util.concurrent.RejectedExecutionException;
import java.util.concurrent.SynchronousQueue;

/**
 * distributes {@code Runnable}s to a fixed number of threads. To keep the
 * code lean, this is not an {@code ExecutorService}. In particular there is
 * only very simple support to shut this executor down.
 */
public class ParallelExecutor implements Executor {
  // other bounded queues work as well and are useful to buffer peak loads
  private final BlockingQueue<Runnable> workQueue =
      new SynchronousQueue<Runnable>();
  private final Thread[] threads;
  
  /*+**********************************************************************/
  /**
   * creates the requested number of threads and starts them to wait for
   * incoming work
   */
  public ParallelExecutor(int numThreads) {
    this.threads = new Thread[numThreads];
    for(int i=0; i<numThreads; i++) {
      // could reuse the same Runner all over, but keep it simple
      Thread t = new Thread(new Runner());
      this.threads[i] = t;
      t.start();
    }
  }
  /*+**********************************************************************/
  /**
   * returns immediately without waiting for the task to be finished, but may
   * block if all worker threads are busy.
   * 
   * @throws RejectedExecutionException if we got interrupted while waiting
   *         for a free worker
   */
  @Override
  public void execute(Runnable task)  {
    try {
      workQueue.put(task);
    } catch (InterruptedException e) {
      Thread.currentThread().interrupt();
      throw new RejectedExecutionException("interrupt while waiting for a free "
          + "worker.", e);
    }
  }
  /*+**********************************************************************/
  /**
   * Interrupts all workers and joins them. Tasks susceptible to an interrupt
   * will preempt their work. Blocks until the last thread surrendered.
   */
  public void interruptAndJoinAll() throws InterruptedException {
    for(Thread t : threads) {
      t.interrupt();
    }
    for(Thread t : threads) {
      t.join();
    }
  }
  /*+**********************************************************************/
  private final class Runner implements Runnable {
    @Override
    public void run() {
      while (!Thread.currentThread().isInterrupted()) {
        Runnable task;
        try {
          task = workQueue.take();
        } catch (InterruptedException e) {
          // canonical handling despite exiting right away
          Thread.currentThread().interrupt(); 
          return;
        }
        try {
          task.run();
        } catch (RuntimeException e) {
          // production code to use a logging framework
          e.printStackTrace();
        }
      }
    }
  }
}

I believe there is quite elegant way to solve this problem by using java.util.concurrent.Semaphore and delegating behavior of Executor.newFixedThreadPool. The new executor service will only execute new task when there is a thread to do so. Blocking is managed by Semaphore with number of permits equal to number of threads. When a task is finished it returns a permit.

public class FixedThreadBlockingExecutorService extends AbstractExecutorService {

private final ExecutorService executor;
private final Semaphore blockExecution;

public FixedThreadBlockingExecutorService(int nTreads) {
    this.executor = Executors.newFixedThreadPool(nTreads);
    blockExecution = new Semaphore(nTreads);
}

@Override
public void shutdown() {
    executor.shutdown();
}

@Override
public List<Runnable> shutdownNow() {
    return executor.shutdownNow();
}

@Override
public boolean isShutdown() {
    return executor.isShutdown();
}

@Override
public boolean isTerminated() {
    return executor.isTerminated();
}

@Override
public boolean awaitTermination(long timeout, TimeUnit unit) throws InterruptedException {
    return executor.awaitTermination(timeout, unit);
}

@Override
public void execute(Runnable command) {
    blockExecution.acquireUninterruptibly();
    executor.execute(() -> {
        try {
            command.run();
        } finally {
            blockExecution.release();
        }
    });
}

I had the same need in the past: a kind of blocking queue with a fixed size for each client backed by a shared thread pool. I ended up writing my own kind of ThreadPoolExecutor:

UserThreadPoolExecutor (blocking queue (per client) + threadpool (shared amongst all clients))

See: https://github.com/d4rxh4wx/UserThreadPoolExecutor

Each UserThreadPoolExecutor is given a maximum number of threads from a shared ThreadPoolExecutor

Each UserThreadPoolExecutor can:

• submit a task to the shared thread pool executor if its quota is not reached. If its quota is reached, the job is queued (non-consumptive blocking waiting for CPU). Once one of its submitted task is completed, the quota is decremented, allowing another task waiting to be submitted to the ThreadPoolExecutor
• wait for the remaining tasks to complete

I found this rejection policy in elastic search client. It blocks caller thread on blocking queue. Code below-

 static class ForceQueuePolicy implements XRejectedExecutionHandler 
 {
        public void rejectedExecution(Runnable r, ThreadPoolExecutor executor) 
        {
            try 
            {
                executor.getQueue().put(r);
            } 
            catch (InterruptedException e) 
            {
                //should never happen since we never wait
                throw new EsRejectedExecutionException(e);
            }
        }
    
        @Override
        public long rejected() 
        {
            return 0;
        }
}

I recently had a need to achieve something similar, but on a ScheduledExecutorService.

I had to also ensure that I handle the delay being passed on the method and ensure that either the task is submitted to execute at the time as the caller expects or just fails thus throwing a RejectedExecutionException.

Other methods from ScheduledThreadPoolExecutor to execute or submit a task internally call #schedule which will still in turn invoke the methods overridden.

import java.util.concurrent.*;

public class BlockingScheduler extends ScheduledThreadPoolExecutor {
    private final Semaphore maxQueueSize;

    public BlockingScheduler(int corePoolSize,
                             ThreadFactory threadFactory,
                             int maxQueueSize) {
        super(corePoolSize, threadFactory, new AbortPolicy());
        this.maxQueueSize = new Semaphore(maxQueueSize);
    }

    @Override
    public ScheduledFuture<?> schedule(Runnable command,
                                       long delay,
                                       TimeUnit unit) {
        final long newDelayInMs = beforeSchedule(command, unit.toMillis(delay));
        return super.schedule(command, newDelayInMs, TimeUnit.MILLISECONDS);
    }

    @Override
    public <V> ScheduledFuture<V> schedule(Callable<V> callable,
                                           long delay,
                                           TimeUnit unit) {
        final long newDelayInMs = beforeSchedule(callable, unit.toMillis(delay));
        return super.schedule(callable, newDelayInMs, TimeUnit.MILLISECONDS);
    }

    @Override
    public ScheduledFuture<?> scheduleAtFixedRate(Runnable command,
                                                  long initialDelay,
                                                  long period,
                                                  TimeUnit unit) {
        final long newDelayInMs = beforeSchedule(command, unit.toMillis(initialDelay));
        return super.scheduleAtFixedRate(command, newDelayInMs, unit.toMillis(period), TimeUnit.MILLISECONDS);
    }

    @Override
    public ScheduledFuture<?> scheduleWithFixedDelay(Runnable command,
                                                     long initialDelay,
                                                     long period,
                                                     TimeUnit unit) {
        final long newDelayInMs = beforeSchedule(command, unit.toMillis(initialDelay));
        return super.scheduleWithFixedDelay(command, newDelayInMs, unit.toMillis(period), TimeUnit.MILLISECONDS);
    }

    @Override
    protected void afterExecute(Runnable runnable, Throwable t) {
        super.afterExecute(runnable, t);
        try {
            if (t == null && runnable instanceof Future<?>) {
                try {
                    ((Future<?>) runnable).get();
                } catch (CancellationException | ExecutionException e) {
                    t = e;
                } catch (InterruptedException ie) {
                    Thread.currentThread().interrupt(); // ignore/reset
                }
            }
            if (t != null) {
                System.err.println(t);
            }
        } finally {
            releaseQueueUsage();
        }
    }

    private long beforeSchedule(Runnable runnable, long delay) {
        try {
            return getQueuePermitAndModifiedDelay(delay);
        } catch (InterruptedException e) {
            getRejectedExecutionHandler().rejectedExecution(runnable, this);
            return 0;
        }
    }

    private long beforeSchedule(Callable callable, long delay) {
        try {
            return getQueuePermitAndModifiedDelay(delay);
        } catch (InterruptedException e) {
            getRejectedExecutionHandler().rejectedExecution(new FutureTask(callable), this);
            return 0;
        }
    }

    private long getQueuePermitAndModifiedDelay(long delay) throws InterruptedException {
        final long beforeAcquireTimeStamp = System.currentTimeMillis();
        maxQueueSize.tryAcquire(delay, TimeUnit.MILLISECONDS);
        final long afterAcquireTimeStamp = System.currentTimeMillis();
        return afterAcquireTimeStamp - beforeAcquireTimeStamp;
    }

    private void releaseQueueUsage() {
        maxQueueSize.release();
    }
}

I have the code here, will appreciate any feedback. https://github.com/AmitabhAwasthi/BlockingScheduler

Here's the solution that seems to work really well. It's called NotifyingBlockingThreadPoolExecutor.

Demo program.

Edit: There is an issue with this code, the await() method is buggy. Calling shutdown() + awaitTermination() seems to work fine.