[Python-Dev] Adding NetworkIOError for bug 1706815

[Guido van Rossum]

On 7/5/07, Gregory P. Smith <greg at electricrain.com> wrote:
> On Wed, Jul 04, 2007 at 11:03:42AM +0200, Guido van Rossum wrote:
> > Why not simply inherit socket.error from EnvironmentError?
>
> True, that would be simpler; is it enough?  If we avoid adding the new
> exception, I really think it should inherit from IOError, not
> EnviromnentError because sockets are I/O.  urllib2.URLError was
> already a child of IOError; doing the same to to socket.error makes
> sense.

OTOH, when os.read() returns an error, os.error (OSError) is raised.
Is that not I/O?

IMO this is all hairsplitting, and the exact inheritance hierarchy for
these doesn't matter much -- nobody is going to write a handler that
treats OSError or socket.error different than IOError.

(If you have different call sites that raise different exceptions,
each call site should have a separate try/except rather than a single
try/except with multiple handlers.)

> The patch makes URLError a child of NetworkIOError instead of IOError.
> Does that make sense?  URLs as an abstract concept may or may not
> imply network I/O behind the scenes though network i/o is the most
> common use.  I could take that argument further and suggest they don't
> necessarily even imply actual I/O if you've provided your own protocol
> handlers.

Well, as long as NetworkIOError inherits from IOError, the change is
compatible, but I don't think anyone would care. Making URLError a
child of EnvironmentError or socket.error could be defended too, and I
doubt it makes a difference for any real code. (Anyone with evidence
to the contrary, please speak up now).

> The question then becomes if there are any use cases for "except
> NetworkIOError:" that code wouldn't want to just use "except IOError:"
> for that using "except socket.error:" or "except urllib2.URLError:"
> are insufficient for.  My intuition is telling me: probably not.
> urllib2 code should catch socket.error everywhere internally and turn
> it into a URLError (the code appears to do this in many places though
> i haven't audited that it does everywhere).

Hm. I'm no fan of such renaming of exceptions (even though the
__cause__ mechanism from PEP 3134 (formerly 344) makes it a little
less problematic).

> -greg
>
> PS for the person complaining that the url didn't work.  blame
>    sourceforge and go look the bug up by id yourself.  nothing i can
>    do about that.

Try python.org/sf/1706815

--Guido

> > On 7/4/07, Gregory P. Smith <greg at electricrain.com> wrote:
> > >In response to bug 1706815 and seeing messy code to catch errors in
> > >network apps I've implemented most of the ideas in the bug and added a
> > >NetworkIOError exception (child of IOError).  With this, socket.error
> > >would now inherit from NetworkIOError instead of being its own thing
> > >(the old one didn't even declare a parent!).
> > >
> > >Complete patch attached to the bug.  All unit tests pass.
> > >Documentation updates included.
> > >
> > > http://sourceforge.net/tracker/index.php?func=detail&aid=1706816&group_id=5470&atid=105470
> > >
> > >Any thoughts?  I'm happy with it and would like to commit it if folks
> > >agree.
>


-- 
--Guido van Rossum (home page: http://www.python.org/~guido/)