Overcoming Software

Dynamic Exception Reporting in Haskell

Exceptions kind of suck in Haskell. You don’t get a stack trace. They don’t show up in the types of functions. They incorporate a subtyping mechanism that feels more like Java casting than typical Haskell programming.

A partial solution to the problem is HasCallStack - that gives us a CallStack which gets attached to error calls. However, it only gets attached to error - so you can either have String error messages and a CallStack, or you can have richly typed exceptions with no location information.

A CallStack is a static piece of information about the code. “You called foo, which called bar, which called quuz, which blew up with Prelude.read: No parse.” The CallStack answers a single question: “Where did this go wrong?”

But there’s often many more interesting questions that simply “Where?” You often want to know Who? When? How? in order to diagnose the big one: why did my code blow up?

In order to help answer these questions and develop robust exception reporting and diagnosing facilities, I created the annotated-exception package.

Better Call Stacks

annotated-exception provides a big improvement in static CallStack behavior. To understand the improvement, let’s dig into the core problem:

Broken Chains and Orphan Stacks

If any function doesn’t include a HasCallStack constraint in your stack, then the chain is broken, and you only get the stack closest to the source.

Consider this trivial example, which has a few ways of blowing up:

import GHC.Stack

foo :: HasCallStack => Int
foo = error "foo"

bar :: HasCallStack => Int
bar = foo

baz :: Int
baz = foo

quux :: HasCallStack => Int
quux = bar

ohno :: HasCallStack => Int
ohno = baz

If we call foo in GHCi, we get the immediate stack trace:

λ> foo
*** Exception: foo
CallStack (from HasCallStack):
  error, called at <interactive>:4:7 in interactive:Ghci1
    foo, called at <interactive>:14:1 in interactive:Ghci2

Since the bar term has the HasCallStack constraint, it will add it’s location to the mix:

λ> bar
*** Exception: foo
CallStack (from HasCallStack):
  error, called at <interactive>:4:7 in interactive:Ghci1
  foo, called at <interactive>:6:7 in interactive:Ghci1
  bar, called at <interactive>:15:1 in interactive:Ghci2

However, baz omits the constraint, which means that you won’t get that function in the stack:

λ> baz
*** Exception: foo
CallStack (from HasCallStack):
  error, called at <interactive>:4:7 in interactive:Ghci1
    foo, called at <interactive>:8:7 in interactive:Ghci1

The quux term has the call stack, so you get the whole story again:

λ> quux
*** Exception: foo
CallStack (from HasCallStack):
  error, called at <interactive>:4:7 in interactive:Ghci1
    foo, called at <interactive>:6:7 in interactive:Ghci1
      bar, called at <interactive>:10:8 in interactive:Ghci1
        quux, called at <interactive>:17:1 in interactive:Ghci2

But here’s the crappy thing - ohno does have a HasCallStack constraint. You might expect that it would show up in the backtrace. But it does not:

λ> ohno
*** Exception: foo
CallStack (from HasCallStack):
  error, called at <interactive>:4:7 in interactive:Ghci1
  foo, called at <interactive>:8:7 in interactive:Ghci1

The CallStack for foo, baz, and ohno are indistinguishable. This makes diagnosing the failure difficult.

To avoid this problem, you must diligently place a HasCallStack constraint on every function in your code base. This is pretty annoying! And if you have any library code that calls your code, the library’s lack of HasCallStack will break your chains for you.

checkpoint to the rescue

annotated-exception introduces the idea of a checkpoint. The simplest one is checkpointCallStack, which attaches the call-site to any exceptions thrown out of the action:

checkpointCallStack
    :: (HasCallStack, MonadCatch m)
    => m a
    -> m a

Let’s replicate the story from above.

import Control.Exception.Annotated

foo :: IO Int
foo = throw (userError "foo")

-- in GHCi, evaluate:
-- λ> foo
*** Exception: 
    AnnotatedException 
         { annotations = 
             [ Annotation @CallStack 
                 [ ( "throw"
                   , SrcLoc 
                         { srcLocPackage = "interactive"
                         , srcLocModule = "Ghci1"
                         , srcLocFile = "<interactive>"
                         , srcLocStartLine = 4
                         , srcLocStartCol = 7
                         , srcLocEndLine = 4
                         , srcLocEndCol = 30
                         }
                   )
                 ]
             ]
         , exception = user error (foo)
         }

I’ve formatted the output to be a bit more legible. Now, instead of a plain IOError, we’ve thrown an AnnotatedException IOError. Inside of it, we have the CallStack from throw, which knows where it was thrown from. That CallStack inside of the exception is reporting the call-site of throw - not the definition site! This is true even though foo does not have a HasCallStack constraint!

Let’s do bar. We’ll do HasCallStack and our checkpointCallStack, just to see what happens:

import GHC.Stack

bar :: HasCallStack => IO Int
bar = checkpointCallStack foo


-- λ> bar
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc { srcLocPackage = "interactive", srcLocModule = "Ghci1", srcLocFile = "<interactive>", srcLocStartLine = 4, srcLocStartCol = 7, srcLocEndLine = 4, srcLocEndCol = 30}
                  )
                , ( "checkpointCallStack"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci2", srcLocFile = "<interactive>", srcLocStartLine = 15, srcLocStartCol = 7, srcLocEndLi ne = 15, srcLocEndCol = 30}
                  )
                , ( "bar"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci3", srcLocFile = "<interactive>", srcLocStartLine = 17, srcLocStartCol = 1, srcLocEndLine = 17, srcLocEndCol = 4}
                  )
                ]
            ]
        , exception = user error (foo)
        }

We get the source location for throw, checkpointCallStack, and then the use site of bar.

Now, suppose we have our Problem Function again: baz doesn’t have a HasCallStack constraint or a checkpointCallStack. And when we called it through ohno, we lost the stack, even though ohno had the HasCallStack constraint.

baz :: IO Int
baz = bar

ohno :: IO Int
ohno = checkpointCallStack baz

-- λ> ohno
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci1", srcLocFile = "<interactive>", srcLocStartLine = 4, srcLocStartCol = 7, srcLocEndLine = 4, srcLocEndCol = 30}
                  )
                , ( "checkpointCallStack"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci2", srcLocFile = "<interactive>", srcLocStartLine = 15, srcLocStartCol = 7, srcLocEndLi ne = 15, srcLocEndCol = 30}
                  )
                , ( "bar"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci3", srcLocFile = "<interactive>", srcLocStartLine = 21, srcLocStartCol = 7, srcLocEndLine = 21, srcLocEndCol = 10}
                  )
                , ( "checkpointCallStack"
                  , SrcLoc {srcLocPackage = "interactive", srcLocModule = "Ghci3", srcLocFile = "<interactive>", srcLocStartLine = 23, srcLocStartCol = 8, srcLocEndLine = 23, srcLocEndCol = 31}
                  )
                ]
            ]
        , exception = user error (foo)
        }

When we call ohno, we preserve all of the entries in the CallStack. checkpointCallStack in ohno adds itself to the CallStack that is present on the AnnotatedException itself, so it doesn’t need to worry about the stack being broken. It’s perfectly capable of recording that history for you.

Ain’t Just a Checkpoint - catch me later

The type signature for catch in annotated-exception looks like this:

catch
    :: (HasCallStack, Exception e, MonadCatch m)
    => m a
    -> (e -> m a)
    -> m a

That HasCallStack constraint is used to give you a CallStack entry for any time that you catch an exception.

newtype MyException = MyException String
    deriving Show

instance Exception MyException

boom :: IO Int
boom = throw (MyException "boom")

recovery :: IO Int
recovery =
    boom `catch` \(MyException message) -> do
        putStrLn message
        throw (MyException (message ++ " recovered"))

recovery catches the MyException from boom, prints the message, and then throws a new exception with a modified message.

λ> recovery
boom
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 19, srcLocStartCol = 9, srcLocEndLine = 19, srcLocEndCol = 54}
                  )
                , ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 13, srcLocStartCol = 8, srcLocEndLine = 13, srcLocEndCol = 34}
                  )
                , ( "catch"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 17, srcLocStartCol = 5, srcLocEndLine = 19, srcLocEndCol = 54}
                  )
                ]
            ]
        , exception = MyException "boom recovered"
        }

Now, look at that call stack: we have the first throw (from boom), then we have the second throw (in recovery), and finally the catch in recovery.

So we know where the exception originally happened, where it was rethrown, and where it was caught. This is fantastic!

But, even better - these annotations survive even if you throw a different type of Exception. This means you can translate exceptions fearlessly, knowing that any essential annotated context won’t be lost.

Dynamic Annotations

As I said earlier, CallStack is fine, but it’s a static thing. We can figure out “what code called what other code” that eventually led to an exception, but we can’t know anything about the running state of the program.

Enter checkpoint. This function attaches an arbitrary Annotation to thrown exceptions. An Annotation is a wrapper around any value that has an instance of Show and Typeable. The library provides an instance of IsString for this, so you can enable OverloadedStrings and have stringly-typed annotations.

constantAnnotation :: IO String
constantAnnotation =
    checkpoint "from constant annotation" $ do
        msg <- getLine
        if null msg
            then throw (MyException "empty message")
            else pure msg

But the real power is in using runtime data to annotate things.

Let’s imagine you’ve got a web application. You’re reporting runtime exceptions to a service, like Bugsnag. Specific teams “own” routes, so if something breaks, you want to alert the right team.

You can annotate thrown exceptions with the route.

data Route 
    = Login
    | Signup
    | ViewPosts
    | CreatePost
    | EditPost PostId
    deriving Show

dispatch :: Request -> IO Response
dispatch req = 
    case parseRequest req of
        Right route ->
            checkpoint (Annotation route) $ 
                case route of
                    Login ->
                        handleLogin
                    Signup -> 
                        handleSignup
                    ViewPosts ->
                        handleViewPosts
                    CreatePost ->
                        handleCreatePost
                    EditPost postId ->
                        checkpoint (Annotation postId) $
                            handleEditPost postId
        Left _ ->
            invalidRouteError

Now, suppose an exception is thrown somewhere in handleLogin. It’s going to bubble up past dispatch and get handled by the Warp default exception handler. That’s going to dig into the [Annotation] and use that to alter the report we send to Bugsnag. The team that is responsible for handleLogin gets a notification that something broke there.

In the EditPost case, we’ve also annotated the exception with the post ID that we’re trying to edit. This means that, when debugging, we can know exactly which post threw the given exception. Now, when diagnosing and debugging, we can immediately pull up the problematic entry. This gives us much more information about the problem, which makes diagnosis easier.

Likewise, suppose we have a function that gives us the logged in user:

withLoggedInUser :: (Maybe (Entity User) -> IO a) -> IO a
withLoggedInUser action = do
    muser <- getLoggedInUser
    checkpoint (Annotation (fmap entityKey muser)) $ do
        action muser

If the action we pass in to withLoggedInUser throws an exception, that exception will carry the Maybe UserId of whoever was logged in. Now, we can easily know who is having a problem on our service, in addition to what the problem actually is.

The Value of Transparency

But wait - if all exceptions are wrapped with this AnnotatedException type, then how do I catch things? Won’t this pollute my codebase?

And, what happens if I try to catch an AnnotatedException MyException but some other code only threw a plain MyException? Won’t that break things?

These are great questions.

catch and try from other libraries will fail to catch a FooException if the real type of the exception is AnnotatedException FooException. However, catch and try from annotated-exception is capable of “seeing through” the AnnotatedException wrapper.

In fact, we took advantage of this earlier - here’s the code for recovery again:

boom :: IO Int
boom = throw (MyException "boom")

recovery :: IO Int
recovery =
    boom `catch` \(MyException message) -> do
        putStrLn message
        throw (MyException (message ++ " recovered"))

Note how catch doesn’t say anything about annotations. We catch a MyException, exactly like you would in Control.Exception, and the annotations are propagated.

But, let’s say you want to catch the AnnotatedException MyException. You just do that.

recoveryAnnotated :: IO Int
recoveryAnnotated =
    boom `catch` \(AnnotatedException annotations (MyException message)) -> do
        putStrLn message
        traverse print annotations
        throw (OtherException (length message))

-- in GHCi,
λ> recoveryAnnotated
boom
Annotation @CallStack [("throw",SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 13, srcLocStartCol = 8, srcLocEndLine = 13, srcLocEndCol = 34})]
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 37, srcLocStartCol = 9, srcLocEndLine = 37, srcLocEndCol = 48}
                  )
                ]
            ]
        , exception = OtherException 4
        }

Now, something tricky occurs here: we don’t preserve the annotations on the thrown exception. If you catch an AnnotatedException, the library assumes that you’re going to handle those yourself.

If you want to keep them, you’d need to throw an AnnotatedException:

recoveryAnnotatedPreserve :: IO Int
recoveryAnnotatedPreserve =
    boom `catch` \(AnnotatedException annotations (MyException message)) -> do
        putStrLn message
        traverse print annotations
        throw (AnnotatedException annotations (OtherException (length message)))

-- in GHCi,
λ> recoveryAnnotatedPreserve 
boom
Annotation @CallStack [("throw",SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 13, srcLocStartCol = 8, srcLocEndLine = 13, srcLocEndCol = 34})]
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 44, srcLocStartCol = 9, srcLocEndLine = 44, srcLocEndCol = 81}
                  )
                ]
            , Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "src/annotated.hs", srcLocStartLine = 13, srcLocStartCol = 8, srcLocEndLine = 13, srcLocEndCol = 34}
                  )
                ]
            ]
        , exception = OtherException 4
        }

We’re missing catch, which is unfortunate, but generally you aren’t going to be doing this - you’re either going to be handling an error completely, or rethrowing it, and the [Annotation] won’t be relevant to you… unless you’re writing an integration with Bugsnag, or reporting on them in some other way.

So annotated-exception’s exception handling functions can “see through” an AnnotatedException inner to work only on the inner exception type. But what if I try to catch a DatabaseException as an AnnotatedException DatabaseException?

Turns out, the Exception instance of AnnotatedException allows you to do that.

import qualified Control.Exception

emptyAnnotationsAreCool :: IO ()
emptyAnnotationsAreCool =
    Control.Exception.throwIO (MyException "definitely not annotated?")
        `Control.Exception.catch`
            \(AnnotatedException annotations (MyException woah)) -> do
                print annotations
                putStrLn woah


-- in GHCi,
λ> emptyAnnotationsAreCool 
[]
definitely not annotated?

We promote the inner into AnnotatedException [] inner. So the library works regardless if any code you throw cares about AnnotatedException. If you call some external library code which throws an exception, you’ll get the first annotation you try - including if that’s just from catch:

catchPutsACallStack :: IO ()
catchPutsACallStack =
    Control.Exception.throwIO (MyException "definitely not annotated?")
        `catch`
            \(MyException woah) -> do
                throw (OtherException (length woah))

-- in GHCi,
λ> catchPutsACallStack 
*** Exception: 
    AnnotatedException 
        { annotations = 
            [ Annotation @CallStack 
                [ ( "throw"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "../parsonsmatt.github.io/src/annotated.hs", srcLocStartLine = 60, srcLocStartCol = 17, srcLocEndLine = 60, srcLocEndCol = 53})
                , ("catch"
                  , SrcLoc {srcLocPackage = "main", srcLocModule = "Annotated", srcLocFile = "../parsonsmatt.github.io/src/annotated.hs", srcLocStartLine = 58, srcLocStartCol = 9, srcLocEndLine = 58, srcLocEndCol = 16}
                  )
                ]
            ]
        , exception = OtherException 25
        }

We get throw and catch both showing up in our stack trace. If we’d used Control.Exception.throwIO instead of Control.Exception.Annotated.throw, then we’d still have catch as an annotation.

Do you feel the power?

The primary purpose here is to share the technique and inspire a hunger for dynamic exception annotations.

We’ve been using this technique at Mercury for most of this year. It has dramatically simplified how we report exceptions, the shape of our exceptions, and how much info we get from a Bugsnag report. It’s now much easier to diagnose problems and fix bugs.

The Really Big Deal here is that - we now have something better than other languages. The lack of stack traces in Haskell is really annoying, and a clear way that Haskell suffers compared to Ruby or Java. But now, with annotated-exception, we actually have more powerful and more useful exception annotations than a mere stack trace. And, since this is all just library functions, you can swap to Control.Exception.Annotated with little fuss.