Pythonic Abstract Base Class / Interface

[Tom Evans]

My basic question:
If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python?

In a staticly typed language like c++ or java, I'd describe the interface
first, then create the classes eithered derived from that ABC or
implementing that interface (same thing really). This was the first
thing I thought to do in python, but then rapidly realised I didn't
actually need to write anything in the base class - as long as code using
one of the implemented classes just tries to do things with it it
would just work. In fact, I didn't even need to make a base class at
all if the calling code never checks for type (as seems normal).

So, as I see it I could:
a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

Tom

 

[Michael Hudson]

My basic question:
If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python? [...]
So, as I see it I could:
a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

I'd say c) was probably the most pythonic, but it depends a bit. a)
might have some documentation value -- for instance, you could
document required relations between the methods in their docstrings.

Cheers,
mwh

 

[Eric Brunel]

My basic question:
If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python?
[...]

So, as I see it I could:
a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

I'd say c) was probably the most pythonic, but it depends a bit. a)
might have some documentation value -- for instance, you could
document required relations between the methods in their docstrings.

Another argument in favor of a) is that pychecker has an option to check if
methods only raising exceptions in a super-class are actually implemented in all
its sub-classes. So:

class MyInterface:
def spam(self):
raise NotImplementedError
class MyClass(MyInterface):
...

has actually more than a documentation value: pychecker should at least report a
warning if the method spam is not defined in MyClass.

HTH

 

[KefX]

So, as I see it I could:

a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

It's funny, I was actually thinking about the same thing not long before this
got posted.

Being a newcomer of sorts to Python, I can't really comment on what's
"Pythonic" or not as well as others can (and only Guido knows for sure), but I
can make some guesses. Pythonic in my mind is clear, concise, but flexible.
This is a case where you can't get all of them at the same time. The most
concise, of course, is answer C, at the possible expense of clarity and
flexibility. Clarity may be more important to you, and the winner for that is
A. (Somebody else suggested using docstrings in the stubs; I second that
opinion.) However, A is less concise.

A and B are both flexible, because you can implement more functions in the base
class, and poof! Derived classes gain the benefits right then and there. But of
course you have to make sure you don't break any of them (if you're careful and
you know what classes are derived from it, and I'm guessing you do, it
shouldn't be a problem). I'm inclined to say that B is more flexible than A,
because if you remove a stub function and some code somewhere relies on it
being called (whether or not it's a no-op), then a derived class that doesn't
define it will be broken. Of course, in practice, this isn't too likely...

To summarize:
C is concise and nothing else
B is concise (not quite as much as C), somewhat clear, flexible
A is clear, flexible (not quite as much as B), but not concise

I'm inclined to pick A, since I'm a bit of a clarity freak, but I don't really
feel that any one of these is more Pythonic than another on its own merits...it
kind of depends on the situation. If subclasses are likely to be small, I would
probably pick C or at most B and save some effort...if subclasses are likely to
be a significant size, I'd definitely pick A.

- Kef

 

[Alex Martelli]

My basic question:
If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python? ...
Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

The normal Python approach is currently to just rely on signature-based
polymorphism -- use inheritance to share implementation, but if you
have no implementation to share, there's no real reason to inherit.

Of course, people coming from languages where inheritance reigns will
not accept that easily and refer to such issues as:

-- in your 'interface class' you can implement the methods as raising
NotImplementedError, which will be clearer than the AttributeError
that happens upon trying to call the method on an instance of a
class that's meant to implement the interface but forgot the
method in question

-- inheritance is ALMOST always centered on sharing implementation --
in the case of exception handling, though, inheritance trees do matter

-- if you inherit you can typetest with isinstance rather than just
featuretest with hasattr (to me, that's a DEFECT of inheritance

If you do have any useful features -- most often Template Method
design patterns -- to share among (some of) your concrete classes,
then by all means use inheritance for that. But, otherwise, I think
that keeping documentation in comments and docstring, and using
other language constructs (including inheritance0 for substantial
purposes and not "just to document things", is a preferable approach.


Alex

 

[Bruno Desthuilliers]

My basic question:
If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python?

In a staticly typed language like c++ or java, I'd describe the interface
first, then create the classes eithered derived from that ABC or
implementing that interface (same thing really). This was the first
thing I thought to do in python, but then rapidly realised I didn't
actually need to write anything in the base class - as long as code using
one of the implemented classes just tries to do things with it it
would just work. In fact, I didn't even need to make a base class at
all if the calling code never checks for type (as seems normal).

So, as I see it I could:
a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

c), unless you have a compelling technical reason -> then a) with
NotImplementedErrors or like... I cannot see what's the b) option would
buy you anyway.

Work with the language, not against it.

Bruno

 

[Fredrik Lundh]

c), unless you have a compelling technical reason -> then a) with
NotImplementedErrors or like... I cannot see what's the b) option would
buy you anyway.

isinstance

</F>

 

[Erik Max Francis]

If I have a specific interface which I know is going to be implemented
by a number of classes, but there is no implementation commonality
between them, what is the preferred form for this in Python? ...
Is there a clear pythonic way, or does it depend a bit more on the
design of the specific program?

I usually do something like

class AbstractBase:
def __init__(self):
if self.__class__ is AbstractBase:
raise NotImplementedError

def aMethod(self):
raise NotImplementedError

def anotherMethod(self):
raise NotImplementedError

 

[Dave Kuhlman]

I usually do something like

class AbstractBase:
def __init__(self):
if self.__class__ is AbstractBase:
raise NotImplementedError

[snip]

If you expect anyone else (or yourself when you've forgotten this
code) to implement another derived class (to implement the
interface), then you will also want to add doc strings to each of
the methods, so that an implementer of your interface knows what
the requirements are for each parameter, return values, etc.
Implementing the (empty) base class is a good way to provide
documentation on your interface.

Dave

 

[Tom Evans]

So, as I see it I could:
a) create a base class with stubs for all the interface and derive from that.
b) create an empty base class and derive from that.
c) just make the concrete classes directly with no inheritance commonality.

Thanks for all the responses. It would appear that there is no
One True Way. What I've learnt:

I was originally just whipping something up, not sure what the final
solution was going to look like. In that case, it would appear (c)
was the way to go.

If I wanted to document exactly what the interface should be once I'm sure
it's the right solution, (a) with document strings and raising
NotImplementedError might be a good place to put that documentation.
(btw, thanks -- I didn't know about NotImplementedError before)

On the face of it, (b) appears to be the loser. But it would give me
isinstance if I wanted it, and might allow me to convert more easily
from (a) to (c) in the future.

thanks to all,
Tom