One-shot timer whose duration can be updated. Think of it as an enhanced version of registerDelay.

This uses GHC.Event when available (GHC 7.2+, -threaded, non-Windows OS). Otherwise, it falls back to forked threads and threadDelay.

Suppose we are managing a network connection, and want to time it out if no messages are received in over five minutes. We'll create a Delay, and an action to "bump" it:

  let timeoutInterval = 5 * 60 * 1000000 :: Int
  delay <- newDelay timeoutInterval
  let bump = updateDelay delay timeoutInterval

This way, the Delay will ring if it is not bumped for longer than five minutes.

Now we fork the receiver thread:

  dead <- newEmptyTMVarIO
  _ <- forkIO $
    (forever $ do
         msg <- recvMessage
         bump
         handleMessage msg
     ) `finally` atomically (putTMVar dead ())

Finally, we wait for the delay to ring, or for the receiver thread to fail due to an exception:

  atomically $ waitDelay delay `orElse` readTMVar dead

Warning:

• If handleMessage blocks, the Delay may ring due to handleMessage taking too long, rather than just recvMessage taking too long.
• The loop will continue to run until you do something to stop it.

It might be simpler to use System.Timeout instead:

  m <- timeout timeoutInterval recvMessage
  case m of
      Nothing  -> fail "timed out"
      Just msg -> handleMessage msg

However, using a Delay has the following advantages:

• If recvMessage makes a blocking FFI call (e.g. network I/O on Windows), timeout won't work, since it uses an asynchronous exception, and FFI calls can't be interrupted with async exceptions. The Delay approach lets you handle the timeout in another thread, while the FFI call is still blocked.
• updateDelay is more efficient than timeout when GHC.Event is available.