On Fri, Oct 15, 2010 at 9:35 AM,
Michael Snoyman wrote:
I have a recommendation of how to fix this: the MonadCatchIO typeclass should be extended to include finally, onException and everything else. We can provide default definitions which will work for most monads, and short-circuiting monads like ErrorT (and I imagine ContT as well) will need to override them.
It seems that `finally' can be fixed without all these proposed additions. The method catch is the only necessary method of the class.
The subject of catching errors in non-IO monads has a long history, some of which is documented at
http://okmij.org/ftp/Haskell/index.html#catch-MonadIO
The page points out to an old CaughtMonadIO file. I have just updated it for new Exceptions, and added the final test:
test3c go = runErrorT $ go `gfinally` (liftIO $ putStrLn "sequel called")
test31 = test3c (return "return" :: ErrorT String IO String)
*CaughtMonadIO> test31 sequel called Right "return"
test32 = test3c (error "error" :: ErrorT String IO String)
*CaughtMonadIO> test32 sequel called *** Exception: error
test33 = test3c (throwError "throwError" :: ErrorT String IO String)
*CaughtMonadIO> test33 sequel called *** Exception: ErrorException "\"throwError\""
As we can see, sequel is always called. Here is the updated file:
http://okmij.org/ftp/Haskell/CaughtMonadIO.lhs
Incidentally, one should be very careful of using `finally' with the continuation monad. The Cont monad lets us ``enter the room once, and exit many times''. So, finally may be called more than once. We need the ugly dynamic-wind -- or try to use less powerful monads if they suffice.
Perhaps I'm misunderstanding your code, but it seems like it's not really respecting the ErrorT monad at all. Instead, it's converting the error type to a runtime exception, which often times is not at all what we want. A pertinent example: in Yesod, I use a modified ErrorT to allow short-circuiting handler functions to perform special responses such as redirects. Using your code, I believe that such code on my part would result in a 500 server error every time, quite the opposite of what I wanted. I would prefer if the test read as:
test33 = fmap (== Left "throwError") $ test3c (throwError "throwError" :: ErrorT String IO String)
Which never in fact returns True. Or, more to the point, the test is never even called, since the runtime exception prevents it. Michael