[Haskell-cafe] Extensible Exceptions

[Duncan Coutts]

On Sun, 2008-11-23 at 01:40 +0100, Henning Thielemann wrote:
> On Sat, 22 Nov 2008, Thomas Schilling wrote:
> 
> > It's a pattern match error, implemented by throwing an asynchronous
> > exception.  The idea being, that we only have one mechanism (well, an
> > synchronous exceptions, thrown via throwIO).
> >
> > Yes, I know that there's a difference between "error" and "exception",
> > but I would argue that which is which depends on the program.  For
> > example, for most programs a pattern match error is a fatal condition,
> > there's no sane recovery from it.  OTOH, in a program like GHCi, a
> > pattern match error in an executed statement is an exceptional
> > condition, which we want to catch, so it doesn't kill GHCi.
> 
> It's completely ok to run something in a sandbox and try to observe 
> errors. But that's debugging and I think there is no need to do this in 
> many places of an application. In general handling errors automatically is 
> not possible, because an error might also be if a program loops 
> infinitely. Thus one should not generally handle an error like an 
> exception.

In general I agree. I would advise against explicitly catching such
exceptions just in the region where one is expecting them. That seems
like bad design.

On the other hand "top level" catch-all handlers that also catch such
logic errors sometimes make sense. For example in a Haskell web server
where we generate a page dynamically it makes a lot of sense to catch
errors in the page-generation function, including pattern match errors,
and produce a 500 error code response and log the error message.

That's a case, rather like ghci, where some flaw in the program can and
should be compartmentalised. There's no attempt to clean up the error
but it is a modular system and there is a clear boundary where failures
can occur without bringing down the entire system.

Duncan