Use threads-supervisor if you want the "poor-man's Erlang supervisors". threads-supervisor is an IO-based library with minimal dependencies which does only one thing: It provides you a Supervisor entity you can use to monitor your forked computations. If one of the managed threads dies, you can decide if and how to restart it. This gives you:

• Protection against silent exceptions which might terminate your workers.
• A simple but powerful way of structure your program into a supervision tree, where the leaves are the worker threads, and the nodes can be other supervisors being monitored.
• A disaster recovery mechanism.

You can install the threads-supervisor library by running: > $ cabal install threads-supervisor

Who worked with Haskell's concurrency primitives would be surely familiar with the forkIO function, which allow us to fork an IO computation in a separate green thread. forkIO is great, but is also very low level, and has a couple of subtleties, as you can read from this passage of the documentation:

The newly created thread has an exception handler that discards the exceptions BlockedIndefinitelyOnMVar,BlockedIndefinitelyOnSTM, and ThreadKilled, and passes all other exceptions to the uncaught exception handler.

To mitigate this, we have a couple of libraries available, for example http://hackage.haskell.org/package/async and http://hackage.haskell.org/package/slave-thread.

But what about if I do not want to take explicit action, but instead specifying upfront how to react to disaster, and leave the library work out the details? This is what this library aims to do.

In this example, let's create four different threads:

job1 :: IO ()
job1 = do
  threadDelay 5000000
  fail "Dead"

This job will die after five seconds.

job2 :: ThreadId -> IO ()
job2 tid = do
  threadDelay 3000000
  killThread tid

This other job instead, we have waited three seconds, and then kill a target thread, generating an asynchronous exception.

job3 :: IO ()
job3 = do
  threadDelay 5000000
  error "Oh boy, I'm good as dead"

This guy is very similar to the first one, except for the fact error is used instead of fail.

job4 :: IO ()
job4 = threadDelay 7000000

job4 is what we wish for all our passing cross computation: smooth sailing.

These jobs represent a significant pool of our everyday computations in the IO monad

Creating a Supervisor is as simple as calling newSupervisor, specifying the RestartStrategy you want to use as well as the size of the EventStream (this depends whether you are using a Bounded supervisor or not). Immediately after doing so, a new thread will be started, monitoring any subsequent IO actions submitted to it.

By default, it's programmer responsibility to read the SupervisionEvent the library writes into its internal queue. If you do not do so, your program might leak. To mitigate this, and to offer a more granular control, two different modules are provided: a Bounded and an Unbounded one, which use, respectively, a TBQueue or a TQueue underneath. You can decide to go with the bounded version, with a queue size enforced by the library author, or pass in your own size.

Let's wrap everything together into a full blown example:

main :: IO ()
main = bracketOnError (do

  sup1 <- newSupervisor OneForOne
  sup2 <- newSupervisor OneForOne

  monitorWith fibonacciRetryPolicy sup1 sup2

  _ <- forkSupervised sup2 fibonacciRetryPolicy job3

  j1 <- forkSupervised sup1 fibonacciRetryPolicy job1
  _ <- forkSupervised sup1 fibonacciRetryPolicy (job2 j1)
  _ <- forkSupervised sup1 fibonacciRetryPolicy job4
  _ <- forkIO (go (eventStream sup1))
  return sup1) shutdownSupervisor (\_ -> threadDelay 10000000000)
  where
   go eS = do
     newE <- atomically $ readTBQueue eS
     print newE
     go eS

What we have done here, was to spawn two supervisors and we have used our swiss knife forkSupervised to spawn four supervised IO computations. As you can see, if we partially apply forkSupervised, its type resemble forkIO one; this is by design, as we want to keep this API as IO-friendly as possible. Note also how we can ask the first supervisor to monitor the second one.

fibonacciRetryPolicy is a constructor for the RetryPolicy, which creates under the hood a RetryPolicy from the "retry" package which is using the fibonacciBackoff. The clear advantage is that you are not obliged to use it if you don't like this sensible default; RetryPolicy is an monoid, so you can compose retry policies as you wish.

The RetryPolicy will also be responsible for determining whether a thread can be restarted or not; in the latter case you will find a ChildRestartedLimitReached in your event log.

If you run this program, hopefully you should see on stdout something like this:

ChildBorn ThreadId 62 2015-02-13 11:51:15.293882 UTC
ChildBorn ThreadId 63 2015-02-13 11:51:15.293897 UTC
ChildBorn ThreadId 64 2015-02-13 11:51:15.293904 UTC
ChildDied ThreadId 61 (MonitoredSupervision ThreadId 61) 2015-02-13 11:51:15.293941 UTC
ChildBorn ThreadId 65 2015-02-13 11:51:15.294014 UTC
ChildFinished ThreadId 64 2015-02-13 11:51:18.294797 UTC
ChildDied ThreadId 63 thread killed 2015-02-13 11:51:18.294909 UTC
ChildDied ThreadId 62 Oh boy, I'm good as dead 2015-02-13 11:51:20.294861 UTC
ChildRestarted ThreadId 62 ThreadId 68 OneForOne 2015-02-13 11:51:20.294861 UTC
ChildFinished ThreadId 65 2015-02-13 11:51:22.296089 UTC
ChildDied ThreadId 68 Oh boy, I'm good as dead 2015-02-13 11:51:25.296189 UTC
ChildRestarted ThreadId 68 ThreadId 69 OneForOne 2015-02-13 11:51:25.296189 UTC
ChildDied ThreadId 69 Oh boy, I'm good as dead 2015-02-13 11:51:30.297464 UTC
ChildRestarted ThreadId 69 ThreadId 70 OneForOne 2015-02-13 11:51:30.297464 UTC
ChildDied ThreadId 70 Oh boy, I'm good as dead 2015-02-13 11:51:35.298123 UTC
ChildRestarted ThreadId 70 ThreadId 71 OneForOne 2015-02-13 11:51:35.298123 UTC

I hope that you are now convinced that this library can be of some use to you!