«

Hi,

I'm implementing Python 3 extension using the Python C API.
I am familiar with defining new types, implementing get/set for attributes, etc.

I'm wondering, is there any mean to implement attribute in module
scope which is read-only?

So, the following

import xyz
print(xyz.flag) # OK
xyz.flag = 0 # error due to no write access

Best regards,
--
Mateusz Loskot, http://mateusz.loskot.net

On Thu, 09 Feb 2012 23:32:59 +1100, Ben Finney wrote:

> Mateusz Loskot <> writes:
>
>> I'm wondering, is there any mean to implement attribute in module scope
>> which is read-only?
>
> Python is designed by and for consenting adults. Rather than
> restricting, instead use conventions to make your intent clear to the
> user of your library.

Oh I agree. The convention I want to use to make my intent clear is the
same convention used for the rest of Python: a runtime exception.

I find this "consenting adults" argument less than convincing. We already
have constants in Python -- every int and float and string is a constant.
You can't modify the object 1 to have the value 42, and for very good
reason, "consenting adults" be damned.

What we don't have is *named* constants.

Python has no shortage of read-only values and members, e.g.:

>>> class A(object):
.... pass
....
>>> A.__dict__ = {}
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: attribute '__dict__' of 'type' objects is not writable


Consider this:

If you pass an int to len(), you get a runtime error telling you that you
did something wrong. Does this violate "consenting adults"? No, if you
want to, you can monkey-patch len to accept ints, but you have to work at
it. The normal behaviour is to get an error, not some arbitrary but
meaningless behaviour. You certainly don't get told to use a naming
convention (Hungarian notation, perhaps) to avoid passing the wrong value
to len().

If you assign to something which is intended to be constant, you don't
get an exception telling you that you made a mistake. Instead, you get
unspecified (but almost certainly incorrect) behaviour in the module, and
told to use a naming convention to remind you not to screw up.

The excuse given is that Python is for "consenting adults", but I don't
believe it. I believe that the real reason is that it is hard to
introduce named constants to Python, and rather than solve that hard
problem, people just whitewash the issue and pretend that it's a feature.
It's not a feature, it's a wart. There is no conceivable use-case for
allowing math.pi = 2.5 to succeed.

Python happily violates "consenting adults" all over the place. We have
properties, which can easily create read-only and write-once attributes.
We have descriptors which can be used for the same. We have immutable
types, and constant values, but not constant names.

Python can enforce all common software contracts I can think of, except
the contract that a name will be set to a specific value. And that is, in
my opinion, a weakness in Python.


--
Steven

On 2/9/2012 8:04 PM, Steven D'Aprano wrote:

> Python happily violates "consenting adults" all over the place. We have
> properties, which can easily create read-only and write-once attributes.

So propose that propery() work at module level, for module attributes,
as well as for class attributes.

--
Terry Jan Reedy

On 10 February 2012 03:27, Terry Reedy <> wrote:
> On 2/9/2012 8:04 PM, Steven D'Aprano wrote:
>
>> Python happily violates "consenting adults" all over the place. We have
>> properties, which can easily create read-only and write-once attributes.
>
>
> So propose that propery() work at module level, for module attributes, as
> well as for class attributes.

I think Steven would like something else: bare names that cannot be
rebound. E.g. something like:

>>> const a = 42
>>> a = 7

Would raise an exception. Is that right?

--
Arnaud

On Thu, 09 Feb 2012 22:27:50 -0500, Terry Reedy wrote:

> On 2/9/2012 8:04 PM, Steven D'Aprano wrote:
>
>> Python happily violates "consenting adults" all over the place. We have
>> properties, which can easily create read-only and write-once
>> attributes.
>
> So propose that propery() work at module level, for module attributes,
> as well as for class attributes.


I'm not wedded to a specific implementation.

Besides, it's not just a matter of saying "property should work in
modules" -- that would require the entire descriptor protocol work for
module lookups, and I don't know how big a can of worms that is. Constant
names is a lot more constrained than computed name lookups.



--
Steven