[Haskell-cafe] thespian thoughts

[Simon Michael]

Hi Alexander,

I went looking for something to clean up rss2irc's thread management, and your recently released thespian package looks like the simplest, most practical actors/erlang/OTP-ish lib for haskell so far. Thanks!

I need to restart threads (actors) in a controlled way when they die or hang. I think some of all of that is also part of Erlang/OTP, and maybe it's already on your todo list. Here are some notes I made, for your interest and in case you have any further thoughts.

I thought of adding some actor-transforming combinators like these:

-- | Spawn the specified action and restart it whenever it dies, unless
--   the restart frequency exceeds any of the given maximum frequencies in
--   which case throw an exception.
restart :: [Frequency] -> IO a -> IO Address
restart fs action = spawn $ do
  setFlag TrapRemoteExceptions
  let start times = []
  forever
   t <- getCurrentTime
   update start times
   when (any fs exceeded by start times) throw error
   worker <- spawn action
   monitor worker
   receive worker [Case _ -> return ()]

-- | Convert hangs to exceptions using a watchdog timer: spawn the
--   specified action under the control of a watcher which monitors it and
--   also kills it and throws an exception if the worker ever fails to
--   send (something) to the watcher within the specified interval.
watchdog :: Seconds -> (Address -> IO a) -> IO Address
watchdog t action = spawn $ do
  watcher <- self
  worker <- spawn $ action watcher
  monitor worker
  forever
   receiveTimeout t [Case _  -> return ()
                     Default -> throw error
                    ]

which I would use something like:

main = do
  (feed, bot) <- ...
  spawn $ do
   reader    <- restart [MaxPerHour 2] $ watchdog 60 $ feedReader feed bot
   announcer <- restart [MaxPerHour 2] $ watchdog 60 $ ircAnnouncer bot
   responder <- restart [MaxPerHour 2] $ watchdog 60 $ ircResponder bot
   monitor reader
   monitor announcer
   monitor responder
   setFlag TrapRemoteExceptions
   receive [Case e :: RemoteException -> exitFailure (show e) -- one of the above died or hung too frequently
           ]

but the types aren't right. I'm not sure if these combinators are possible with the current thespian api. Also they would complicate things, eg what is the address of a restarted actor ?

On second thoughts it might be better if these features were built in to the library. I imagine something like:

main = do ...
  reader    <- spawnWith [RestartSpec [MinInterval (minutes 5), MaxPerHour 3, MaxPerWeek 6], Watchdog (seconds 60)] $ feedReader feed bot

feedReader = do
  forever
    poll feed, do stuff
    resetWatchdog

What do you think ?

Best - Simon
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/haskell-cafe/attachments/20111009/b9d9b1b5/attachment.htm>