Beyond Exception Handling: Conditions and Restarts | Hacker News Reader

Beyond Exception Handling: Conditions and Restarts

gigamonkeys.com

103 pointsDanielRibeiro20 comments

This is the one feature I wish I could steal from Common Lisp in almost every language I work with. It would be much easier to put in than almost any other big lisp feature, and damn would it make a lot of things easier.

This is something you can do in most languages if all the code that needs restarts follows a convention/protocol. Red Daly implemented CL condition handling for JavaScript for the Parenscript CL-to-JS compiler: https://github.com/gonzojive/paren-psos/commit/6578ad223515d...

The compiled CL code knows about the convention and follows it, but other JavaScript code doesn't and isn't affected by it.

Yeah, that's why I talk about stealing it at the language level. Although it's possible to use this based on conventions, the real power comes from being able to use it to cover weird conditions while gluing various libraries together. If your language already has a relatively similar tool (i.e. "traditional" exceptions) there's basically no chance that anything you pick up will support conditions.

That said, it's a wonderful model to know about/follow if you come into the rare case that makes reimplementing it where you can useful.

This is true for Common Lisp libraries as well - if they don't offer restarts and just throw exceptions, it looks like it does in any other language. Developing a good protocol for signalling and handling errors takes work.

Is there any other language that implements this? I'll freely admit that I'm no expert on Lisp, so I've tried digging through the code, and I sorta get what the concept is, but I don't really understand it. Any other languages have the same idea?

Dylan does, and my language Slate (http://www.slatelanguage.org) does, as well as its related younger siblings, Atomo (http://atomo-lang.org) and Atomy (http://atomy-lang.org). And all of these implement them in the context of an object model, rather than carefully avoiding CLOS dependencies as in CL.

I just added conditions & restarts today to Fancy (http://fancy-lang.org/). It's currently in a branch but it works =).

Thanks for doing that. More implementations means there's a wider spectrum of places for developers to learn these tools and maybe even improve on them.

I believe Mesa did. On the other hand, Stroustrup claims (in the Design and Evolution of C++, I think) that the Cedar system written in Mesa basically never used the ability to restart as a reason why C++ doesn't have continuable exceptions. I've never been able to track down more information about that than his claim though.

Kent Pitman wrote a pretty thorough history: http://www.nhplace.com/kent/Papers/Condition-Handling-2001.h...

Here's a good explanation using pseudo-Java: http://axisofeval.blogspot.com/2011/04/whats-condition-syste...

Does this basically mean that CL provides a mechanism that's lower level than what is otherwise equivalent to exceptions/warnings, and that this in turn enables things like restarts? Is this analogous to the relationship between higher level control structures such as if/for/while/break/continue vs the lower-level goto? Lastly, can some of these features not be implemented with exceptions?:

  def parse_log_file(f):
      result = []
      for line in f.readlines(): # Fails for large files
        parsed_line = None
        retries = 0
        while retries < 5:
            try:
                result.append(parse_log_entry(line))
                break
            except MalformedLogEntry:
                line = fix(line)
                retries += 1
      return result

Your code is not equivalent. There is a separation of concerns that is possible with conditions/restarts that is not possible with exceptions.

The calling code can choose the appropriate restarts base on information that the callee code does not - and should not - have access to.

> The calling code can choose the appropriate restarts base on information that the callee code does not - and should not - have access to.

In the common case, is that not equivalent to the caller catching specific errors by wrapping the callee in a try/except? Could you elaborate?

But this is assuming the caller has enough information to solve the problem, which is not always the case (the callee could have information on the structure of the disk that are not relevant to the caller for instance).

Conditions/restarts make a separation of concerns, particularly when the caller knows which of the restart is the more appropriate, if the callee has not enough information/context to choose, and if the callee has more information on how to solve the problem.

I guess the separation can be summarized by: difference between knowing WHAT and HOW to do something. A useful abstraction I guess.

I see. So this would be somewhat like saying

  except MalformedLogEntry, ex:
      line = fix(line, ex)

and having ex encode the information that fix() needs to do its job?

If that's the case, then I can see the value in having the ability to pass such information around baked into the language. The code I have is growing more and more ugly, and the only way to simplify it is to introduce some kind of a coding convention that is not common to Python.

How would you do, in your case, if the parse_log_file function does not know the best strategy if a log file is corrupted? Maybe this information is known only by the caller.

I don't know how to solve that in a language without restarts. The information of what to do could well be asked to the final user (with a dialog box) once an error is catched, the callee would then restart at the point where the execution was interrupted applying the chosen strategy.

EDIT: rereading your Python code, I guess you're right and have the closest equivalent.

Ah. So the point is that the state of execution of parse_log_entry and everything up the stack from it is frozen, and then an out-of-band exception handler figures out what to do, fixes the condition that caused the error, and resumes from the exact spot where the error was first detected, restarting just the most low-level code block? Man, CL is pretty sweet.

This, by no coincidence I'm sure, reminds me a lot of how signals are handled in UNIX/POSIX. I guess you could simulate this very crudely with those by sending yourself a SIGUSR[1|2] :).

Maybe I misunderstand it, but doesn't "retries" gets increased only on MalformedLogEntries? I mean, if there's no malformed log entries -- how will it ever exit?

Fixed. I changed how the data is appended to the result half way through writing it so there was no condition to break out of the loop. Now there is a break after the append.

Reply on news.ycombinator.com