On Sat, Feb 9, 2013 at 11:31 PM, R. David Murray wrote:

The reason we avoid such type checks is that we prefer to operate via "duck typing", which means that if an object behaves like the expected input, it is accepted. Here, if we did an explicit type check for str, it would prevent join from working on an "act alike" object that had just enough str like methods to work correctly in os.join (for example, some specialized object that was among other things a filename proxy). Granted, this is pretty far-fetched for os.path.join, but the general principle applies :)

I occasionally ponder the possibility of a time where we have a string.Text ABC, and can stop hardcoding checks for isinstance(str) everywhere :P (No, I'm not volunteering to write it. And if anyone does want to discuss it further, please start a new thread on python-ideas, as python-dev isn't the right venue) Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia

On Sat, 09 Feb 2013 14:35:33 +0000, Thomas Scrace wrote:

R. David Murray writes:

...

The reason we avoid such type checks is that we prefer to operate via "duck typing", which means that if an object behaves like the expected input, it is accepted. Here, if we did an explicit type check for str, it would prevent join from working on an "act alike" object that had just enough str like methods to work correctly in os.join (for example, some specialized object that was among other things a filename proxy).

I see, that makes sense. Thanks. I guess this actually goes to the heart of the flexibility of dynamic/weakly-typed languages. If we wanted to strictly enforce the type of a function's arguments we would use a strong type system.

No, it is more the difference between *statically* typed and dynamically typed. Python is a strongly typed language (every object has a specific type). --David

On 2/9/2013 8:31 AM, R. David Murray wrote:

On Sat, 09 Feb 2013 09:59:13 +0000, Thomas Scrace wrote:

...

If a function (or other non-string object) is accidentally passed as an argument to os.path.join() the result is an AttributeError:

In [3]: os.path.join(fn, "path")

...

--------------------------------------------------------------------------- AttributeError Traceback (most recent call last) /home/tom/<ipython-input-3-44b097ceab04> in <module>() ----> 1 os.path.join(fn, "path") /usr/lib/python2.7/posixpath.pyc in join(a, *p) 66 if b.startswith('/'): 67 path = b ---> 68 elif path == '' or path.endswith('/'): 69 path += b 70 else: AttributeError: 'function' object has no attribute 'endswith'

It's relatively easy to run into this if you mean to pass the return value of a function (fn()) as the argument but accidentally forget to append parens (()) to the callable, thus passing the function itself instead.

Would it not be more helpful to raise a TypeError("Argument must be a string") than the slightly more mysterious AttributeError?

It's not the most difficult error in the world to figure out, I admit, but a TypeError just seems like the more correct thing to do here.

I agree. Since the exception type is not documented and since no one should intentionally pass anything but strings, and therefore should not be writing try: os.path.join(a,b) except AttributeError: barf() I think it would be acceptable to make a change in 3.4.

We don't generally do that kind of type checking in Python. Sometimes the error that results is obscure enough that adding an early check of some sort is worth it, but this one is very clear, so I don't see that an additional check would add much value here.

Changing AttributeError to TypeError only requires try-except, which is cheap if there is no error, not an early type check.

The reason we avoid such type checks is that we prefer to operate via "duck typing", which means that if an object behaves like the expected input, it is accepted.

And catching the existing AttributeError allows work-alikes. I agree that an isinstance check is bad as well as unnecessary. As it is, the core code for path, given above, is *already* wrapped in try: except TypeError, (to give a more friendly error message!, as TypeError is most like from a unicode-bytes mixing). So there would be no extra cost for correct calls and all that is needed is another except clause. except AttributeError as e: bad = e.args[0].split()[0] msg = "all components must be strings; one is a {}".format(bad) raise TypeError("all components must be strings") Thomas, unless David or someone else shoots down this idea here, I say go ahead and open an enhancement issue and add terry.reedy as nosy. -- Terry Jan Reedy

On 9 February 2013 20:08, R. David Murray wrote:

Especially given that there is already a try/except there, this seems fine to me. I think perhaps the error text would be better if it referred only to the type that is invalid, not to str. So something like:

TypeError("object of type {} is not valid as a path "component".format(type(bad)))

--David

I agree that this is a better error message. I have raised an enhancement issue as Terry suggested: http://bugs.python.org/issue17174

On 2/9/2013 6:23 PM, Greg Ewing wrote:

Terry Reedy wrote:

...

I agree. Since the exception type is not documented and since no one should intentionally pass anything but strings, and therefore should not be writing

try: os.path.join(a,b) except AttributeError: barf()

I think it would be acceptable to make a change in 3.4.

Why should we go out of our way to address this particular case, and not any of the infinitely many other situations where you could forget to add parens to a function call?

I agree we do not want to go down this slippery slope. The number of such places in the standard library is huge. The existing error message is one that a user should be able to understand, since it's so common. -- Eric.