2010/8/5 Fred Drake <fdrake@acm.org>:

2010/8/4 Łukasz Langa <lukasz@langa.pl>:

1. The patch makes KeyError behave analogically to IOError so that the first

arg is now a message and the second is the actual key.

I agree with Antoine; there's no point to this.

2. Some people suggest adding e.key to KeyError. I like the idea but in my

opinion currently it is not implementable in a reliable way.

This is interesting and useful.

I'd be really happy to see e.key be present if the key is known

(because it was specifically provided to the constructor:

KeyError(key=...)), or not present if the key isn't known.  (The idea

is much less interesting if code can't distinguish between the

key-is-known and the key-not-known cases.)

The runtime and standard library should be adjusted to provide the key

whenever possible, of course.

Though I doubt this would break anything, we've lived without this

long enough that the it doesn't represent a sufficient failing that

the moratorium should be broken.  It can wait.

+1 on what Fred said (i.e. post-moratorium, add a keyword-only "key"
argument to KeyError, set "e.key" only if that argument is supplied,
update the standard library to supply it and use a default message of
"'Key not found: %r' % key" if the key argument is supplied without an
explicit message). Also +1 for doing the equivalent with
AttributeError and an "attr" keyword only argument.

Since a keyword-only approach doesn't actually *break* any current
code, I'm only -0 on doing that for 3.2 rather than -1.