commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gilles Sadowski <>
Subject Re: [math] right way to throw IndexOutOfBoundsException?
Date Mon, 08 Aug 2011 21:01:04 GMT
On Mon, Aug 08, 2011 at 09:55:49PM +0200, Luc Maisonobe wrote:
> Le 08/08/2011 21:25, Phil Steitz a écrit :
> >On 8/8/11 12:03 PM, Gilles Sadowski wrote:
> >>Hi.
> >>
> >>>>[...]
> >>>>>>Cases such as this would fall in the "illegal argument" category.
> >>>>>>Thus:
> >>>>>>
> >>>>>>   throw new OutOfRangeException(index, 0, parameters.length);
> >>>>>>
> >>>>>>or, to get a more detailed message,
> >>>>>>
> >>>>>>   OutOfRangeException e = new OutOfRangeException(index, 0,
> >>>>>>   e.addMessage(INDEX, index);
> >>>>>>   throw e;
> >>>>>>
> >>>>>>
> >>>>>>Of course, "OutOfRangeException" cannot inherit from both
> >>>>>>"IllegalArgumentException" and "IndexOutOfBoundsException"...
> >>>>>I thought about that, but would prefer to throw
> >>>>>ArrayIndexOutOfBoundsException because that is really what is going
> >>>>>on and I would prefer to throw the standard exception.  Ideally,
> >>>>>would like to throw that with a message reporting the value and the
> >>>>>length of the array.  So, there are three choices:
> >>>>>
> >>>>>0) throw AIOB with no message
> >>>>>1) subclass and throw with localized message - my suggestion above
> >>>>>2) OutOfRangeException
> >>>>>
> >>>>>I like 1) the best and since we have decided to deprecate the
> >>>>>MathRuntimeException, note that it applies to all of the other
> >>>>>standard exceptions generated by MathRuntimeException's createXxx
> >>>>>methods that have not yet been subclassed.  I think we should follow
> >>>>>the generally accepted practice to favor standard exceptions, so
> >>>>>that means we are going to have to create wrappers for all that we
> >>>>>use.  I am willing to help with this.  In this case, I will go ahead
> >>>>>and add the MathArrayIndexOutOfBoundsException that will be an
> >>>>>ArrayIndexOutOfBoundsException if others are OK with this.   Note
> >>>>>that doing this will allow us to handle situations where IAE is not
> >>>>>appropriate (essentially why AIOB does not itself extend IAE).
> >>>>As I understand it, AIOB is a low-level exception that is thrown by the
> >>>>checking an array access:
> >>>>---CUT---
> >>>>   i = 3;
> >>>>   double a = arr[i]; //<--- can throw AIOB
> >>>>---CUT---
> >>>>
> >>>>However, I don't see how a user code can be similar to this: When you
> >>>>the index in your above code, you didn't try to access the array yet.
> >>>>detected that it won't work because the index value is "out of range",
> >>>>"illegal".
> >>>When we are about to access an array, we can perform the check.  We
> >>>can either just allow the JVM to throw the RTE (essentially my
> >>>option 0) above) or provide some more context info to the user (my
> >>>option 1) or throw an entirely different exception (option 2).  It
> >>>is good to throw some kind of AIOB when that is in fact what is
> >>>going on, as it provides more context info than just "something is
> >>>out of range" (OutOfRangeException).
> >>As shown above, you can add as many items of context information as you want
> >>with the exception context. The example I've suggested above will create a
> >>message that will print something like:
> >>
> >>   OutOfRangeException: 3 out of [0, 2] range: index (3)
> >>
> >>I find this quite clear; but we can even add another "LocalizedFormats" like
> >>"ARRAY_INDEX" if you really need the above to read:
> >>
> >>   OutOfRangeException: 3 out of [0, 2] range: array index (3)
> >>
> >>However, the crux of my point is that the "array" part is an implementation
> >>detail. A user should not care that a sequence of data is stored in an array
> >>of primitives or in a "List" or a CM's "RealVector". The real info is that
> >>the index used to access the data must fall within a range; if not, it is
> >>"out of range".
> >>
> >>Then, what I was saying in the previous post is that we should not throw
> >>AIOB because we are not the JVM. In Java, that exception is a *result* of
> >>"calling" the [] operator. Your test happens before calling it; and an
> >>"OutOfRangeException" is as accurate as necessary but not more (so as not to
> >>leak about the internals of a class).
> >
> >I can see that we are not going to agree on this.  I prefer standard
> >exceptions and it is generally accepted best practice to favor
> >them.  To say that only the JDK should throw JDK RTEs is silly,
> >IMO.  We are going to have the same problem with ConcurrentMod, IO,
> >EOF and the others currently created by MathRuntimeException.  Are
> >we going to try to shoehorn every use of any of these into some kind
> >of "some number too small" or other special [math] exceptions?  What
> >do others think about this?
> Well, I basically gave up on exceptions :-(
> Here are my views, but I would not enforce them anymore on anybody.
> I consider JDK exceptions could be used, and that we needed the
> localization (I know the view of everyone on this, so don't argue
> again). this was the reason for the createXxx exception, which had
> signatures corresponding to standard exceptions, and did subclass
> them with anonymous classes that override the JDK standard
> getMessage and getLocalizedMessage methods. This way, we combined
> both needs.

At the cost of strange-looking code:
  throw new MathRuntime.createXxxException("some message", args);
instead of the cleaner:
  throw new XxxException(args);

All needs (detailed info + localization) are still covered by the "new"
(standard) way: This was the compromise we agreed on. But it is perpetually
being eroded!

> Once again, this are only my own views. Do what you want with them.

All I have asked for is to let a chance to the new design. If users complain
at the release of 3.0, then fine you'll go back to the old days, but it's
not fair to act as if nothing had been done for over one year.


To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message