That's the kind of thing I had in mind, but I was also thinking of
subtyping-like relations in the exception hierarchy. For example, does
it make sense to change a contract to raise a less specific error? Does
it matter?
At any rate, it seems we might need code to be able to catch and handle
different contract violations different ways. Raising only one kind of
error is fine when humans are handling them, but programs like Plot are
too stupid to read error messages.
From your response, it seems safe for now to insert another level in
the exception hierarchy for math domain errors, so I'll make a bug
report about it and assign it to myself.
Neil ⊥
On 05/21/2014 12:29 PM, Robby Findler wrote:
Oh! Interesting thought. Right now, the contract system always raises
the same exception.
But I'm having trouble seeing what's the generalization.
It would be easy to make contracts that do the same checking as
existing contracts but that raise more specific errors, and even have
combinators that take a "normal" contract and turn it into one that
raises a specific exception, ie a contract that checks to see if its
input is a number and not positive and raises one exception in that
cases and a different exception in the case that it gets a non-number
entirely.
Is that what you have in mind?
Robby
On Wed, May 21, 2014 at 1:17 PM, Neil Toronto <neil.toro...@gmail.com> wrote:
I thought there might be a contract-system-y reason; e.g. if I wanted to
change `gamma` to raise errors on exact, nonpositive integers using a
contract. Or maybe something deeper that has to do with composing contracts
using `or/c` and `and/c`. Nothing like that?
Neil ⊥
On 05/21/2014 12:09 PM, Robby Findler wrote:
I don't think there is a problem with adding that exception, except
that it seems pretty laborious. And probably someone (and I can't
think of anyone better than you to do this, Neil!) may find that there
are actually 2 or 3 exns that are the right set.
I think that if you wanted to take this on, Racketeers everywhere
would thank you.
Robby
On Wed, May 21, 2014 at 1:05 PM, Neil Toronto <neil.toro...@gmail.com>
wrote:
Based on experience using Plot, I believe allowing all errors is a
terrible
default.
Jay's suggestion would make it easier to ignore only some errors,
especially
if it were a parameter. I could see doing
(plot-swallow-exception? exn:fail:contract:divide-by-zero?)
To make it as backwards-compatible as possible while being precise, it
would
be nice if it could swallow only math domain errors by default. But
again,
there's no way to distinguish those from other errors.
Is there a problem with adding `exn:fail:contract:math` and making
`exn:fail:contract:divide-by-zero` a subtype?
Neil ⊥
On 05/21/2014 11:25 AM, Jay McCarthy wrote:
Or
#:catch-these-exns-plot! [catch-exn? void]
Jay
On Wed, May 21, 2014 at 11:22 AM, Robby Findler
<ro...@eecs.northwestern.edu> wrote:
How about adding yet another argument to plot #:errors-go-to-nan or
something like that, such that plot's default behavior is to let the
errors thru, but that, if you turn that flag on, then it just swallows
all errors?
Robby
On Wed, May 21, 2014 at 12:18 PM, Neil Toronto <neil.toro...@gmail.com>
wrote:
Division by zero and other math domain errors, primarily. There are
two
good
reasons to ignore these.
1. Plotting is for visualizing functions. Users shouldn't have to
ensure
to
Plot that the function they're plotting is total when all they want is
to
*see* it, not use it for some critical calculation.
More concretely, raising all exceptions means that this:
(plot3d (surface3d / -1 1 -1 1))
has to be this instead:
(plot3d (surface3d
(λ (x y) (with-handlers ([exn? (λ (e) +nan.0)])
(/ x y)))
-1 1 -1 1))
or it has to use an explicit domain test (i.e. test (= y 0)). That's a
lot
of work just to visualize the division function.
2. Plotting necessarily approximates. Plot's behavior should be as
consistent as possible regardless of exactly how it approximates.
Currently, plotting the division function in any subdomain or with any
sampling interval just works. Suppose Plot allowed errors to be
raised.
Then
this would plot just fine:
(plot3d (surface3d / -2 1 -2 1))
because sampling 41 points (the default) within [-2,1] skips over 0.
But
this would raise an error:
(plot3d (list (surface3d / -2 1 -2 1)
(surface3d + -1 2 -1 2)))
because the bounds are extended to [-2,2]x[-2,2] accommodate the `+`
surface
renderer, so 0 would be sampled.
----
In a sense, what I really want is impossible. With any
exception-swallowing
rule I make, someone could write a function that raises an exception
that
Plot eats when it shouldn't, or re-raises when it shouldn't. Someone
could
try plotting a function that, for some reason, tries to compute (gamma
0) if
a `vector-ref` goes out of bounds.
But I think it could behave properly in the most common cases if it
were
possible to tell the difference between the errors raised by (gamma 0)
and
(vector-ref vs (length vs)). Right now, there's no way to do that.
Neil ⊥
On 05/21/2014 10:16 AM, Robby Findler wrote:
I meant to suggest that plot should raise any error that a function
it
calls raises. Is there something wrong with doing that?
Robby
On Wed, May 21, 2014 at 11:09 AM, Neil Toronto
<neil.toro...@gmail.com>
wrote:
Referencing an identifier before its definition raises an
`exn:fail:contract:variable`. So that would put Plot in the weird
position
of having to distinguish different kinds of contract errors, but not
by
using subtypes or any other simple rule.
Maybe we should have an `exn:fail:contract:math` or
`exn:fail:contract:domain` or `exn:fail:contract:argument` and make
`exn:fail:contract:divide-by-zero` a subtype of it. I'd gladly
change
Plot
to ignore just those, and change the math library to raise more
specific
errors.
(log 0) raises an `exn:fail:contract:divide-by-zero`, which is a
terminology
stretch that suggests we need a bit of refactoring anyway.
Neil ⊥
On 05/21/2014 09:51 AM, Robby Findler wrote:
IMO, plot should raise that exception.
Robby
On Wed, May 21, 2014 at 10:37 AM, Laurent
<laurent.ors...@gmail.com>
wrote:
... it eats everything! (particularly exceptions)
For example, it is perfectly happy with the following:
% racket
Welcome to Racket v6.0.1.7.
(require plot)
(plot (function (lambda(x)(+ x n)))
#:x-min 0 #:x-max 10 #:y-min 0 #:y-max 10)
(object:2d-plot-snip% ...)
... although n is clearly undefined. Presumably this is to avoid
breaking
on
math errors like `(/ 0)` ?
Would it make sense to have an `exn:fail:arithmetic` exception
struct
so
that `function` would only catch those?
Laurent
____________________
Racket Users list:
http://lists.racket-lang.org/users
____________________
Racket Users list:
http://lists.racket-lang.org/users
____________________
Racket Users list:
http://lists.racket-lang.org/users
____________________
Racket Users list:
http://lists.racket-lang.org/users
____________________
Racket Users list:
http://lists.racket-lang.org/users
