Core Java

Asynchronous retry pattern

When you have a piece of code that often fails and must be retried, this Java 7/8 library provides rich and unobtrusive API with fast and scalable solution to this problem:

ScheduledExecutorService scheduler = Executors.newSingleThreadScheduledExecutor();
RetryExecutor executor = new AsyncRetryExecutor(scheduler).
    withExponentialBackoff(500, 2).     //500ms times 2 after each retry
    withMaxDelay(10_000).               //10 seconds
    withUniformJitter().                //add between +/- 100 ms randomly

You can now run arbitrary block of code and the library will retry it for you in case it throws SocketException:

final CompletableFuture<Socket> future = executor.getWithRetry(() ->
        new Socket("localhost", 8080)

future.thenAccept(socket ->
        System.out.println("Connected! " + socket)

Please look carefully! getWithRetry() does not block. It returns CompletableFuture immediately and invokes given function asynchronously. You can listen for that Future or even for multiple futures at once and do other work in the meantime. So what this code does is: trying to connect to localhost:8080 and if it fails with SocketException it will retry after 500 milliseconds (with some random jitter), doubling delay after each retry, but not above 10 seconds.

Equivalent but more concise syntax:

        getWithRetry(() -> new Socket("localhost", 8080)).
        thenAccept(socket -> System.out.println("Connected! " + socket));

This is a sample output that you might expect:

TRACE | Retry 0 failed after 3ms, scheduled next retry in 508ms (Sun Jul 21 21:01:12 CEST 2013) Connection refused
    at Method) ~[na:1.8.0-ea]

TRACE | Retry 1 failed after 0ms, scheduled next retry in 934ms (Sun Jul 21 21:01:13 CEST 2013) Connection refused
    at Method) ~[na:1.8.0-ea]

TRACE | Retry 2 failed after 0ms, scheduled next retry in 1919ms (Sun Jul 21 21:01:15 CEST 2013) Connection refused
    at Method) ~[na:1.8.0-ea]

TRACE | Successful after 2 retries, took 0ms and returned: Socket[addr=localhost/,port=8080,localport=46332]

Connected! Socket[addr=localhost/,port=8080,localport=46332]

Imagine you connect to two different systems, one is slow, second unreliable and fails often:

CompletableFuture<String> stringFuture = executor.getWithRetry(ctx -> unreliable());
CompletableFuture<Integer> intFuture = executor.getWithRetry(ctx -> slow());

stringFuture.thenAcceptBoth(intFuture, (String s, Integer i) -> {
    //both done after some retries

thenAcceptBoth() callback is executed asynchronously when both slow and unreliable systems finally reply without any failure. Similarly (using CompletableFuture.acceptEither()) you can call two or more unreliable servers asynchronously at the same time and be notified when the first one succeeds after some number of retries.

I can’t emphasize this enough – retries are executed asynchronously and effectively use thread pool, rather than sleeping blindly.


Often we are forced to retry given piece of code because it failed and we must try again, typically with a small delay to spare CPU. This requirement is quite common and there are few ready-made generic implementations with retry support in Spring Batch through RetryTemplate class being best known. But there are few other, quite similar approaches ([1], [2]). All of these attempts (and I bet many of you implemented similar tool yourself!) suffer the same issue – they are blocking, thus wasting a lot of resources and not scaling well.

This is not bad per se because it makes programming model much simpler – the library takes care of retrying and you simply have to wait for return value longer than usual. But not only it creates leaky abstraction (method that is typically very fast suddenly becomes slow due to retries and delay), but also wastes valuable threads since such facility will spend most of the time sleeping between retries. Therefore
Async-Retry utility was created, targeting Java 8 (with Java 7 backport existing) and addressing issues above.

The main abstraction is RetryExecutor that provides simple API:

public interface RetryExecutor {

    CompletableFuture<Void> doWithRetry(RetryRunnable action);

    <V> CompletableFuture<V> getWithRetry(Callable<V> task);

    <V> CompletableFuture<V> getWithRetry(RetryCallable<V> task);

    <V> CompletableFuture<V> getFutureWithRetry(RetryCallable<CompletableFuture<V>> task);

Don’t worry about RetryRunnable and RetryCallable – they allow checked exceptions for your convenience and most of the time we will use lambda expressions anyway.

Please note that it returns CompletableFuture. We no longer pretend that calling faulty method is fast. If the library encounters an exception it will retry our block of code with preconfigured backoff delays. The invocation time will sky-rocket from milliseconds to several seconds. CompletableFuture clearly indicates that. Moreover it’s not a dumb java.util.concurrent.Future we all know – CompletableFuture in Java 8 is very powerful and most importantly – non-blocking by default.

If you need blocking result after all, just call .get() on Future object.

Basic API

The API is very simple. You provide a block of code and the library will run it multiple times until it returns normally rather than throwing an exception. It may also put configurable delays between retries:

RetryExecutor executor = //...

executor.getWithRetry(() -> new Socket("localhost", 8080));

Returned CompletableFuture<Socket> will be resolved once connecting to localhost:8080 succeeds. Optionally we can consume RetryContext to get extra context like which retry is currently being executed:

    getWithRetry(ctx -> new Socket("localhost", 8080 + ctx.getRetryCount())).

This code is more clever than it looks. During first execution ctx.getRetryCount() returns 0, therefore we try to connect to localhost:8080. If this fails, next retry will try localhost:8081 (8080 + 1) and so on. And if you realize that all of this happens asynchronously you can scan ports of several machines and be notified about first responding port on each host:

Arrays.asList("host-one", "host-two", "host-three").
    forEach(host ->
            getWithRetry(ctx -> new Socket(host, 8080 + ctx.getRetryCount())).

For each host RetryExecutor will attempt to connect to port 8080 and retry with higher ports.

getFutureWithRetry() requires special attention. I you want to retry method that already returns CompletableFuture<V>: e.g. result of asynchronous HTTP call:

private CompletableFuture<String> asyncHttp(URL url) { /*...*/}


final CompletableFuture<CompletableFuture<String>> response = 
    executor.getWithRetry(ctx -> 
        asyncHttp(new URL("")));

Passing asyncHttp() to getWithRetry() will yield CompletableFuture<CompletableFuture<V>>. Not only it’s awkward to work with, but also broken. The library will barely call asyncHttp() and retry only if it fails, but not if returned
CompletableFuture<String> fails. The solution is simple:

final CompletableFuture<String> response =
    executor.getFutureWithRetry(ctx ->
        asyncHttp(new URL("")));

In this case RetryExecutor will understand that whatever was returned from asyncHttp() is the actually just a Future and will (asynchronously) wait for result or failure. This library is much more powerful, so let’s dive into:

Configuration options

In general there are two important factors you can configure: RetryPolicy that controls whether next retry attempt should be made and Backoff – that optionally adds delay between subsequent retry attempts.

By default RetryExecutor repeats user task infinitely on every Throwable and adds 1 second delay between retry attempts.

Creating an instance of RetryExecutor

Default implementation of RetryExecutor is AsyncRetryExecutor which you can create directly:

ScheduledExecutorService scheduler = Executors.newSingleThreadScheduledExecutor();

RetryExecutor executor = new AsyncRetryExecutor(scheduler);



The only required dependency is standard ScheduledExecutorService from JDK. One thread is enough in many cases but if you want to concurrently handle retries of hundreds or more tasks, consider increasing the pool size.

Notice that the AsyncRetryExecutor does not take care of shutting down the ScheduledExecutorService. This is a conscious design decision which will be explained later.

AsyncRetryExecutor has few other constructors but most of the time altering the behaviour of this class is most convenient with calling chained with*() methods. You will see plenty of examples written this way. Later on we will simply use executor reference without defining it. Assume it’s of RetryExecutor type.

Retrying policy


By default every Throwable (except special AbortRetryException) thrown from user task causes retry. Obviously this is configurable. For example in JPA you may want to retry a transaction that failed due to OptimisticLockException – but every other exception should fail immediately:

    getWithRetry(ctx -> dao.optimistic());

Where dao.optimistic() may throw OptimisticLockException. In that case you probably don’t want any delay between retries, more on that later. If you don’t like the default of retrying on every Throwable, just restrict that using retryOn():


Of course the opposite might also be desired – to abort retrying and fail immediately in case of certain exception being thrown rather than retrying. It’s that simple:

        getWithRetry(ctx -> dao.optimistic());

Clearly you don’t want to retry NullPointerException or IllegalArgumentException as they indicate programming bug rather than transient failure. And finally you can combine both retry and abort policies. User code will retry in case of any retryOn() exception (or subclass) unless it should abortOn() specified exception. For example we want to retry every IOException or SQLException but abort if FileNotFoundException or java.sql.DataTruncation is encountered (order is irrelevant):

    getWithRetry(ctx -> dao.load(42));

If this is not enough you can provide custom predicate that will be invoked on each failure:

    abortIf(throwable ->
        throwable instanceof SQLException &&

Max number of retries

Another way of interrupting retrying “loop” (remember that this process is asynchronous, there is no blocking loop) is by specifying maximum number of retries:


In rare cases you may want to disable retries and barely take advantage from asynchronous execution. In that case try:


Delays between retries (backoff)

Retrying immediately after failure is sometimes desired (see OptimisticLockException example) but in most cases it’s a bad idea. If you can’t connect to external system, waiting a little bit before next attempt sounds reasonably. You save CPU, bandwidth and other server’s resources. But there are quite a few options to consider:

  • should we retry with constant intervals or increase delay after each failure?
  • should there be a lower and upper limit on waiting time?
  • should we add random “jitter” to delay times to spread retries of many tasks in time?

This library answers all these questions.

Fixed interval between retries

By default each retry is preceded by 1 second waiting time. So if initial attempt fails, first retry will be executed after 1 second. Of course we can change that default, e.g. to 200 milliseconds:


If we are already here, by default backoff is applied after executing user task. If user task itself consumes some time, retries will be less frequent. For example with retry delay of 200ms and average time it takes before user task fails at about 50ms RetryExecutor will retry about 4 times per second (50ms + 200ms). However if you want to keep retry frequency at more predictable level you can use fixedRate flag:


This is similar to “fixed rate” vs. “fixed delay” approaches in ScheduledExecutorService. BTW don’t expect RetryExecutor to be very precise, it does it’s best but it heavily depends on aforementioned ScheduledExecutorService accuracy.

Exponentially growing intervals between retries

It’s probably an active research subject, but in general you may wish to expand retry delay over time, assuming that if the user task fails several times we should try less frequently. For example let’s say we start with 100ms delay until first retry attempt is made but if that one fails as well, we should wait two times more (200ms). And later 400ms, 800ms… You get the idea:

executor.withExponentialBackoff(100, 2)

This is an exponential function that can grow very fast. Thus it’s useful to set maximum backoff time at some reasonable level, e.g. 10 seconds:

    withExponentialBackoff(100, 2).
    withMaxDelay(10_000)      //10 seconds

Random jitter

One phenomena often observed during major outages is that systems tend to synchronize. Imagine a busy system that suddenly stops responding. Hundreds or thousands of requests fail and are retried. It depends on your backoff but by default all these requests will retry exactly after one second producing huge wave of traffic at one point in time. Finally such failures are propagated to other systems that, in turn, synchronize as well.

To avoid this problem it’s useful to spread retries over time, flattening the load. A simple solution is to add random jitter to delay time so that not all request are scheduled for retry at the exact same time. You have choice between uniform jitter (random value from -100ms to 100ms):

executor.withUniformJitter(100)     //ms

…and proportional jitter, multiplying delay time by random factor, by default between 0.9 and 1.1 (10%):

executor.withProportionalJitter(0.1)        //10%

You may also put hard lower limit on delay time to avoid to short retry times being scheduled:

executor.withMinDelay(50)   //ms

Implementation details

This library was built with Java 8 in mind to take advantage of lambdas and new CompletableFuture abstraction (but Java 7 port with Guava dependency exists). It uses ScheduledExecutorService underneath to run tasks and schedule retries in the future – which allows best thread utilization.

But what is really interesting is that the whole library is fully immutable, there is no single mutable field, at all. This might be counter-intuitive at first, take for example this trivial code sample:

ScheduledExecutorService scheduler = Executors.newSingleThreadScheduledExecutor();

AsyncRetryExecutor first = new AsyncRetryExecutor(scheduler).
    withExponentialBackoff(500, 2);

AsyncRetryExecutor second = first.abortOn(FileNotFoundException.class);

AsyncRetryExecutor third = second.withMaxRetries(10);

It might seem that all with*() methods or retryOn()/ abortOn() mutate existing executor. But that’s not the case, each configuration change creates new instance, leaving the old one untouched. So for example while first executor will retry on FileNotFoundException, the second and third won’t. However they all share the same scheduler. This is the reason why AsyncRetryExecutor does not shut down ScheduledExecutorService (it doesn’t even have any close() method). Since we have no idea how many copies of AsyncRetryExecutor exist pointing to the same scheduler, we don’t even try to manage its lifecycle. However this is typically not a problem (see Spring integration below).

You might be wondering, why such an awkward design decision? There are three reasons:

  • when writing a concurrent code immutability can greatly reduce risk of multi-threading bugs. For example RetryContext holds number of retries. But instead of mutating it we simply create new instance (copy) with incremented but final counter. No race condition or visibility can ever occur.
  • if you are given an existing RetryExecutor which is almost exactly what you want but you need one minor tweak, you simply call executor.with...() and get a fresh copy. You don’t have to worry about other places where the same executor was used (see: Spring integration for further examples)
  • functional programming and immutable data structures are sexy these days.

N.B.: AsyncRetryExecutor is not marked final, does you can break immutability by subclassing it and adding mutable state. Please don’t do this, subclassing is only permitted to alter behaviour.


This library requires Java 8 and SLF4J for logging. Java 7 port additionally depends on Guava.

Spring integration

If you are just about to use RetryExecutor in Spring – feel free, but the configuration API might not work for you. Spring promotes (or used to promote) the convention of mutable services with plenty of setters. In XML you define bean and invoke setters (via <property name="..."/>) on it. This convention assumes the existence of mutating setters. But I found this approach error-prone and counter-intuitive under some circumstances.

Let’s say we globally defined bean and injected it in multiple places. Great. Now there is this one single request that requires slightly different timeout:

private TransactionTemplate template;

and later in the same class:

final int oldTimeout = template.getTimeout();
//do the work

This code is wrong on so many levels! First of all if something fails we never restore oldTimeout. OK, finally to the rescue. But also notice how we changed global, shared TransactionTemplate instance. Who knows how many other beans and threads are just about to use it, unaware of changed configuration?

And even if you do want to globally change the transaction timeout, fair enough, but it’s still wrong way to do this. private timeout field is not volatile and thus changes made to it may or may not be visible to other threads. What a mess! The same problem appears with many other classes like JmsTemplate.

You see where I’m going? Just create one, immutable service class and safely adjust it by creating copies whenever you need it. And using such services is equally simple these days:

class Beans {

    public RetryExecutor retryExecutor() {
        return new AsyncRetryExecutor(scheduler()).
            withExponentialBackoff(500, 2);

    @Bean(destroyMethod = "shutdownNow")
    public ScheduledExecutorService scheduler() {
        return Executors.newSingleThreadScheduledExecutor();


Hey! It’s 21st century, we don’t need XML in Spring any more. Bootstrap is simple as well:

final ApplicationContext context = new AnnotationConfigApplicationContext(Beans.class);
final RetryExecutor executor = context.getBean(RetryExecutor.class);

As you can see integrating modern, immutable services with Spring is just as simple. BTW if you are not prepared for such a big change when designing your own services, at least consider constructor injection.


This library is covered with a strong battery of unit tests. However it wasn’t yet used in any production code and the API is subject to change. Of course you are encouraged to submit bugs, feature requests and pull requests. It was developed with Java 8 in mind but Java 7 backport exists with slightly more verbose API and mandatory Guava dependency ( ListenableFuture instead of
CompletableFuture from Java 8).

Full source code on GitHub.

Reference: Asynchronous retry pattern from our JCG partner Tomasz Nurkiewicz at the Java and neighbourhood blog.

Tomasz Nurkiewicz

Java EE developer, Scala enthusiast. Enjoying data analysis and visualization. Strongly believes in the power of testing and automation.
Notify of

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Newest Most Voted
Inline Feedbacks
View all comments
Tomer Ben David
Tomer Ben David
10 years ago


You are an amazing developer. I want to be like you. How do you manage to have the time for these great blogs and open sources?

6 years ago

Excelent article on the things I wanted to implement.

Back to top button