Safe, consistent, and easy exception handling
Runtime exceptions - as exposed in
Control.Exceptionmodule - have long been an intimidating part of the Haskell ecosystem. This package, and this README for the package, are intended to overcome this. By providing an API that encourages best practices, and explaining the corner cases clearly, the hope is to turn what was previously something scary into an aspect of Haskell everyone feels safe using.
This is an initial release of the package. I fully expect the library to expand in the near future, and in particular there are two open issues for decisions that need to be made in the short term. I'm releasing the package in its current state since:
On the second point, I've created a
to get feedback on the interruptible/uninterruptible issue and the
naming issue. Both are described in this blog post.
I'm hoping this library can bring some sanity and comfort to people dealing
IO and wanting to ensure proper exception handling! Following is the
content of the README, which can also be read on
This package provides additional safety and simplicity versus
Control.Exception by having its functions recognize the difference between
synchronous and asynchronous exceptions. As described below, synchronous
exceptions are treated as recoverable, allowing you to catch and handle them
as well as clean up after them, whereas asynchronous exceptions can only be
cleaned up after. In particular, this library prevents you from making the
This section is intended to give you the bare minimum information to use this library (and Haskell runtime exceptions in general) correctly.
Control.Exception.Safemodule. Do not import
Control.Exceptionitself, which lacks the safety guarantees that this library adds. Same applies to
throw. (For compatible naming, there are synonyms for this of
If you need to perform some allocation or cleanup of resources, use
one of the following functions (and don't use the
try family of functions):
Hopefully this will be able to get you up-and-running quickly.
Request to readers: if there are specific workflows that you're unsure of how to accomplish with this library, please ask so we can develop a more full-fledged cookbook as a companion to this file.
We're going to define three different versions of exceptions. Note that these definitions are based on how the exception is thrown, not based on what the exception itself is:
Synchronous exceptions are generated by the current thread. What's important about these is that we generally want to be able to recover from them. For example, if you try to read from a file, and the file doesn't exist, you may wish to use some default value instead of having your program exit, or perhaps prompt the user for a different file location.
Asynchronous exceptions are thrown by either a different user
thread, or by the runtime system itself. For example, in the
race will kill the longer-running thread with
an asynchronous exception. Similarly, the
timeout function will
kill an action which has run for too long. And the runtime system
will kill threads which appear to be deadlocked on
In contrast to synchronous exceptions, we almost never want to recover from asynchronous exceptions. In fact, this is a common mistake in Haskell code, and from what I've seen has been the largest source of confusion and concern amongst users when it comes to Haskell's runtime exception system.
Impure exceptions are hidden inside a pure value, and exposed
by forcing evaluation of that value. Examples are
impureThrow. Additionally, incomplete pattern
matches can generate impure exceptions. Ultimately, when these
pure values are forced and the exception is exposed, it is thrown
as a synchronous exception.
Since they are ultimately thrown as synchronous exceptions, when it comes to handling them, we want to treat them in all ways like synchronous exceptions. Based on the comments above, that means we want to be able to recover from impure exceptions.
If we never want to be able to recover from asynchronous exceptions,
why do we want to be able to catch them at all? The answer is for
resource cleanup. For both sync and async exceptions, we would like
to be able to acquire resources - like file descriptors - and register
a cleanup function which is guaranteed to be run. This is exemplified
by functions like
So to summarize:
Unfortunately, GHC's runtime system provides no way to determine if an exception was thrown synchronously or asynchronously, but this information is vitally important. There are two general approaches to dealing with this:
Run an action in a separate thread, don't give that thread's ID to anyone else, and assume that any exception that kills it is a synchronous exception. This approach is covered in the School of Haskell article catching all exceptions, and is provided by the enclosed-exceptions package.
Make assumptions based on the type of an exception, assuming that certain exception types are only thrown synchronously and certain only asynchronously.
Both of these approaches have downsides. For the downsides of the type-based approach, see the caveats section at the end. The problems with the first are more interesting to us here:
IO, it gets trickier for pure functions and monad transformer stacks. The latter issue is solved via monad-control and the exceptions packages. The former issue, however, means that it's impossible to provide a universal interface for failure for pure and impure actions. This may seem esoteric, and if so, don't worry about it too much.
Therefore, this package takes the approach of trusting type
information to determine if an exception is asynchronous or
synchronous. The details are less interesting to a user, but the
basics are: we leverage the extensible extension system in GHC and
state that any extension type which is a child of
is an async exception. All other exception types are assumed to be
Once we're able to distinguish between sync and async exceptions, and we know our goals with sync vs async, how we handle things is pretty straightforward:
finally), we don't care if the exception is sync or async: call the cleanup function and then rethrow the exception.
With this explanation, it's useful to consider async exceptions as "stronger" or more severe than sync exceptions, as the next section will demonstrate.
One annoying corner case is: what happens if, when running a cleanup function after an exception was thrown, the cleanup function itself throws an exception. For this, we'll consider
action `onException` cleanup. There are four different possibilities:
Our guiding principle is: we cannot hide a more severe exception with
a less severe exception. For example, if
action threw a sync
exception, and then
cleanup threw an async exception, it would be a
mistake to rethrow the sync exception thrown by
action, since it
would allow the user to recover when that is not desired.
Therefore, this library will always throw an async exception if either the action or cleanup thows an async exception. Other than that, the behavior is currently undefined as to which of the two exceptions will be thrown. The library reserves the right to throw away either of the two thrown exceptions, or generate a new exception value completely.
The exceptions package provides an abstraction for throwing, catching, and cleaning up from exceptions for many different monads. This library leverages those type classes to generalize our functions.
There are a few choices of naming that differ from the base libraries:
throwin this library is for synchronously throwing within a monad, as opposed to in base where
throwIOserves this purpose and
throwis for impure throwing. This library provides
impureThrowfor the latter case, and also provides convenience synonyms
catchfunction in this package will not catch async exceptions. Please use
catchAsyncif you really want to catch those, though it's usually better to use a function like
withExceptionwhich ensure that the thrown exception is rethrown.
Let's talk about the caveats to keep in mind when using this library.
There is a big debate and difference of opinion regarding checked versus unchecked exceptions. With checked exceptions, a function states explicitly exactly what kinds of exceptions it can throw. With unchecked exceptions, it simply says "I can throw some kind of exception." Java is probably the most famous example of a checked exception system, with many other languages (including C#, Python, and Ruby) having unchecked exceptions.
As usual, Haskell makes this interesting. Runtime exceptions are most
assuredly unchecked: all exceptions are converted to
Exception typeclass, and function signatures do not state
which specific exception types can be thrown (for more on this, see
next caveat). Instead, this information is relegated to documentation,
and unfortunately is often not even covered there.
By contrast, approaches like
EitherT are very explicit
in the type of exceptions that can be thrown. The cost of this is that
there is extra overhead necessary to work with functions that can
return different types of exceptions, usually by wrapping all possible
exceptions in a sum type.
This library isn't meant to settle the debate on checked vs unchecked, but rather to bring sanity to Haskell's runtime exception system. As such, this library is decidedly in the unchecked exception camp, purely by virtue of the fact that the underlying mechanism is as well.
Another advantage of the
EitherT approach is that you are
explicit in your function signature that a function may fail. However,
the reality of Haskell's standard libraries are that many, if not the
vast majority, of
IO actions can throw some kind of exception. In
fact, once async exceptions are considered, every
IO action can
throw an exception.
Once again, this library deals with the status quo of runtime
exceptions being ubiquitous, and gives the rule: you should consider
IO type as meaning both that a function modifies the outside
world, and may throw an exception (and, based on the previous
caveat, may throw any type of exception it feels like).
There are attempts at alternative approaches here, such as unexceptionalio. Again, this library isn't making a value statement on one approach versus another, but rather trying to make today's runtime exceptions in Haskell better.
As explained above, this library makes heavy usage of type information
to differentiate between sync and async exceptions. While the approach
used is fairly well respected in the Haskell ecosystem today, it's
certainly not universal, and definitely not enforced by the
Control.Exception module. In particular,
throwIO will allow you to
synchronously throw an exception with an asynchronous type, and
throwTo will allow you to asynchronously throw an exception with a
The functions in this library prevent that from happening via
exception type wrappers, but if an underlying library does something
surprising, the functions here may not work correctly. Further, even
when using this library, you may be surprised by the fact that
Foo `catch` (\Foo -> ...) won't actually trigger the exception
Foo looks like an asynchronous exception.
The ideal solution is to make a stronger distinction in the core libraries themselves between sync and async exceptions.
Two exceptions types which are handled surprisingly are
though these exceptions are thrown asynchronously by the runtime
system, for our purposes we treat them as synchronous. The reasons are
takeMVaris synchronously throwing
This discussion is now being tracked at: https://github.com/fpco/safe-exceptions/issues/3
Control.Exception, allocation functions and cleanup handlers in
bracket are masked using the (interruptible)
function, in contrast to
uninterruptibleMask. There have been some debates
about the correctness of this in the past, notably a libraries mailing list
discussion kicked off by Eyal
It seems that general consensus is:
uninterruptibleMaskis a better choice
In its current version, this library uses
mask (interruptible) for allocation
uninterruptibleMask cleanup handlers. This is a debatable
decision (and one worth debating!). An example of alternatives would be:
uninterruptibleMaskfor both allocation and cleanup pieces
We may decide to rename
throw to something else at some point. Please see