--------------------------------------------------------------------------- CHAPTER 4. THE SECRETS OF MULTIPLE INHERITANCE --------------------------------------------------------------------------- Introduction --------------------------------------------------------------------------- In my opinion, Multiple Inheritance (often abbreviated as MI) is the most advanced topic in Object Oriented Programming. It is also true that Multiple Inheritance is one of the most difficult features to implement in an Object Oriented Programming language. Even, some languages by design decided to avoid it. This is for instance the case of Java, that avoided MI having seen its implementation in C++ (which is not for the faint of heart ;-); for what concerns the so called scripting languages, of which the most famous are Perl, Python and Ruby (in this order, even if the right order would be Python, Ruby and Perl), only Python implements Multiple Inheritance. The fact that Multiple Inheritance can be hairy, does not mean that it is *always* hairy, however. Multiple Inheritance is used with success in Lisp derived languages (CLOS, Dylan,..), for instance. The aims of this chapter is to show how much the Python support for MI has improved in the most recent version (2.2 and 2.3). The message is the following: if Python 1.5 had a basic support for MI inheritance (basic but nevertheless with nice features, dynamic), Python 2.2 has *greatly* improved that support and with the change of the Method Resolution Order in Python 2.3, we may say that support for MI is now *excellent*. I strongly encourage Python programmers to use MI a lot: this will allows even a stronger reuse of code than in single inheritance. Why Multiple Inheritance ? -------------------------- I think the best way to explain the power of MI is trough a simple example. #singleton, noninstantiable The Python 2.3 Method Resolution Order -------------------------------------- *Felix qui potuit de rerum cognoscere causas* -- Virgilius Everything started with a post by Samuele Pedroni in the Python development mailing list [[#]_]. In his post, Samuele showed that the Python 2.2 method resolution order is not monotonic and he proposed to replace it with the C3 method resolution order. Guido agreed with his arguments and therefore now Python 2.3 uses C3. The C3 method itself has nothing to do with Python, since it has been invented by people working on Dylan and it is described in a paper intended for lisper readers [[#]_]. The present paper gives a (hopefully) readable discussion of the C3 algorithm for Pythonistas who want to understand the reasons for the change. First of all, let me point out that all I am going to say only applies to the *new style classes* introduced in Python 2.2: *classic classes* maintain their old method resolution order, depth first and then left to right. Therefore, there is no breaking of old code for classic classes; and even if in principle there could be breaking of code for Python 2.2 new style classes, in practice the cases in which the C3 resolution order differs from the Python 2.2 method resolution order are so rare that no real breaking of Python 2.2 code is expected. Therefore: *don't be scared !* Moreover, unless you make strong use of multiple inheritance and you have non-trivial hierarchies, you don't need to understand the C3 algorithm, and you can easily skip this paper. On the other hand, if you really want to knows how multiple inheritance works, then this paper is for you. The good news is that things are not as complicated as you could expect. Let me begin with some basic definitions. 1) Given a class in a complicate multiple inheritance hierarchy, it is a non-trivial task to specify the order in which methods are overridden, i.e. to specify the order of the ancestors of the class. 2) The list of the ancestors of a class C, including the class itself, ordered from the nearest ancestor to the furthest, is called the class precedence list or the *linearization* of the class C. 3) The *Method Resolution Order* (MRO) is the set of rules that allow to construct the linearization. In the Python literature, the idiom "the MRO of C" is also used as a synonymous for the linearization of the class C. 4) In the case of a single inheritance hierarchy in which C is a subclass of C1 which in turn is a subclass of C2, the linearization of C is the list [C, C1 , C2]. In multiple inheritance hierarchies, the construction of the linearization is cumbersome, since it has to respect the essential constraints of *local precedence ordering* and *monotonicity*. 5) I will discuss the local precedence ordering later, but I can give the definition of monotonicity here. A MRO is monotonic when the following it true: *if C1 precedes C2 in the linearization of C, then C1 precedes C2 in the linearization of any subclass of C*. Otherwise, the innocuous operation of deriving a new class could change the resolution order of methods, potentially introducing very subtle bugs. Examples where this happens will be shown later. 6) Not all classes admit a linearization. There are cases, in complicated hierarchies, where it is not possible to derive a class such that its linearization respects all the good properties. Here I give an example of this situation. Consider the hierarchy :: O = object class X(O): pass class Y(O): pass class A(X,Y): pass class B(Y,X): pass which can be represented with the following inheritance graph, where I have denoted with O the class object, which is the beginning of any hierarchy for new style classes: :: ----------- | | | O | | / \ | - X Y / | / | / | / |/ A B \ / ? In this case, it is not possible to derive a new class C from A and B, since X precedes Y in A, but Y precedes X in B, therefore the method resolution order would be ambiguous in C. Python 2.3 raises an exception in this situation (TypeError: MRO conflict among bases Y, X) forbidding the naive programmer from creating ambiguous hierarchies. Python 2.2 instead does not raise an exception, but chooses an *ad hoc* ordering (ZABXYO in this case). The C3 Method Resolution Order ------------------------------ Let me introduce few simple notations which will be useful for the following discussion. I use the shortcut notation C1 C2 ... CN to indicate the list of classes [C1, C2, ... , CN]. The *head* of the list is its first element: head = C1 whereas the *tail* is the rest of the list: tail = C2 ... CN. I also use the notation C + (C1 C2 ... CN) = C C1 C2 ... CN to denote the sum of the lists [C] + [C1, C2, ... ,CN]. Now I can explain how the MRO works in Python 2.3, i.e. I give the rules to compute the linearization of any given class C. Rule 0: If C is the "object" class, which has no parents, its linearization coincides with itself: L(O) = O Rule 1: If C = C(B) is a class with an unique parent B (single inheritance) its linearization is simply the sum of the class with the linearization of the parent: L(C(B)) = C + L(B) Rule 2: In the case of multiple inheritance things are more cumbersome and one has to introduce the concept of *merge* of lists. Consider for instance the case of two parents B1 and B2. Then the linearization of C is given by C plus the merge of the three sequences: - linearization of the first parent; - linearization of the second parent; - list of the two parents In symbolic notation, L(C(B1,B2)) = C + merge(L(B1), L(B2), B1 B2) How the merge is computed ? The rule is the following: *take the head of the first sequence, i.e L(B1)[0]; if this head is not in the tail of any of the other sequences, then add it to the linearization of C and remove it from the sequences in the merge, otherwise look at the head of the next sequence and take it, if it is a good head. Then repeat the operation. If there are no good heads, then it is impossible to construct the merge. In this case Python 2.3 will raise an error and will refuse to create the class C.* This prescription ensures that the merge operation *preserves* the ordering, if the ordering can be preserved. On the other hand, if the order cannot be preserved (as in the example of serious order disagreement discussed before) then the merge cannnot be computed. The generalization to the case of a class with three or more parents is obvious: L(C(B1, ... , BN)) = C + merge(L(B1), ... ,L(BN), B1 ... BN) The merge rule is the heart of the C3 algorithm: once you have understood that, you have understood the C3 Method Resolution Order. But I don't expect that you can understand the rule without a couple of examples ;-) Examples -------- First example. Consider the following hierarchy: >>> O = object >>> class F(O): pass >>> class E(O): pass >>> class D(O): pass >>> class C(D,F): pass >>> class B(D,E): pass >>> class A(B,C): pass In this case the inheritance graph can be drawn as :: 6 --- Level 3 | O | (more general) / --- \ / | \ | / | \ | / | \ | --- --- --- | Level 2 3 | D | 4| E | | F | 5 | --- --- --- | \ \ _ / | | \ / \ _ | | \ / \ | | --- --- | Level 1 1 | B | | C | 2 | --- --- | \ / | \ / \ / --- Level 0 0 | A | (more specialized) --- The linearizations of O,D,E and F are trivial:: L(O) = O L(D) = D O L(E) = E O L(F) = F O The linearization of B can be computed as L(B) = B + merge(DO, EO, DE) We see that D is a good head, therefore we take it and we are reduced to compute merge(O,EO,E). Now O is not a good head, since it is in the tail of the sequence EO. In this case the rule says that we have to skip to the next sequence. Then we see that E is a good head; we take it and we are reduced to compute merge(O,O) which gives O. Therefore L(B) = B D E O With the same argument one finds:: L(C) = C + merge(DO,FO,DF) = C + D + merge(O,FO,F) = C + D + F + merge(O,O) = C D F O Now we can compute :: L(A) = A + merge(BDEO,CDFO,BC) = A + B + merge(DEO,CDFO,C) = A + B + C + merge(DEO,DFO) = A + B + C + D + merge(EO,FO) = A + B + C + D + E + merge(O,FO) = A + B + C + D + E + F + merge(O,O) = A B C D E F O In this example, the linearization is ordered in a pretty nice way according to the inheritance level, in the sense that lower levels (i.e. more specialized classes) have higher precedence (see the inheritance graph). However, this is not the general case. I leave as an exercise for the reader to compute the linearization for my second example: >>> O = object >>> class F(O): pass >>> class E(O): pass >>> class D(O): pass >>> class C(D,F): pass >>> class B(E,D): pass >>> class A(B,C): pass The only difference with the previous example is the change B(D,E) --> B(E,D); however even such a little modification completely changes the ordering of the hierarchy:: 6 --- Level 3 | O | / --- \ / | \ / | \ / | \ --- --- --- Level 2 2 | E | 4 | D | | F | 5 --- --- --- \ / \ / \ / \ / \ / \ / --- --- Level 1 1 | B | | C | 3 --- --- \ / \ / --- Level 0 0 | A | --- Notice that the class E, which is in the second level of the hierarchy, precedes the class C, which is in the first level of the hierarchy, i.e. E is more specialized than C, even if it is in an upper level. The lazy programmer can obtain the MRO directly from Python 2.2, since in this case it coincides with the Python 2.3 linearization. It is enough to invoke the .mro() method of class A: >>> A.mro() (, , , , , , ) Finally, let me consider the example discussed in the first section, involving a serious order disagreement. In this case, it is obvious to compute the linearizations of the classes O, X, Y, A and B:: L(O) = 0 L(X) = X O L(Y) = Y O L(A) = A X Y O L(B) = B Y X O However, it is impossible to compute the linearization for a class C that inherits from A and B:: L(C) = C + merge(AXYO, BYXO, AB) = C + A + merge(XYO, BYXO, B) = C + A + B + merge(XYO, YXO) At this point we cannot merge the lists XYO and YXO, since X is in the tail of YXO whereas Y is in the tail of XYO: therefore there are no good heads and the C3 algorithm stops. Python 2.3 raises an error and refuses to create the class C. Bad Method Resolution Orders ---------------------------- A MRO is *bad* when it breaks such fundamental properties as local precedence ordering and/or monotonicity. In this section, I will show that both the MRO for classic classes and the MRO for new style classes in Python 2.2 are bad. It is easier to start with the local precedence ordering. Consider the following example: >>> F=type('Food',(),{remember2buy:'spam'}) >>> E=type('Eggs',(F,),{remember2buy:'eggs'}) >>> G=type('GoodFood',(F,E),{}) We see that the class G inherits from F and E, with F *before* E: therefore we would expect the attribute *G.remember2buy* to be inherited by *F.rembermer2buy* and not by *E.remember2buy*: nevertheless Python 2.2 gives >>> G.remember2buy 'eggs' This is a breaking of local precedence ordering since the order in the local precedence list, i.e. the list of the parents of G, is not preserved in the Python 2.2 linearization of G: L(G,P22)= G E F object # F *follows* E One could argue that the reason why F follows E in the Python 2.2 linearization is that F is less specialized than E, since F is the superclass of E; nevertheless the breaking of local precedence ordering is quite non-intuitive and error prone. This is particularly true since there is a difference with old style classes: >>> class F: remember2buy='spam' >>> class E(F): remember2buy='eggs' >>> class G(F,E): pass >>> G.remember2buy 'spam' In this case the MRO is GFEF and the local precedence ordering is preserved. As a general rule, hierarchies such as :: O | F | \ | E | / G should be avoided, since it is unclear if F should override E or viceversa. Python 2.3 solves the ambiguity buy raising an exception in the creation of class G, effectively stopping the programming from generating ambiguous hierarchies. The reason for that is that the C3 algorithm fails, since the merge merge(FO,EFO,FE) cannot be computed, because F is in the tail of EFO and E is in the tail of FE. The real solution is to design a non-ambiguous hierarchy, i.e. to derive G from E and F (the more specific first) and not from F and E; in this case the MRO is GEF without any doubt. Python 2.3 forces the programmer to write good hierarchies (or, at least, less prone to errors). On a related note, let me point out that the Python 2.3 algorithm is smart enough to recognize obvious mistakes, as the duplication of classes in the list of parents: >>> class A(object): pass >>> class C(A,A): pass Traceback (most recent call last): File "", line 1, in ? TypeError: duplicate base class A Python 2.2 (both for classic classes and new style classes) in this situation would not raise any error message. Finally, I would like to point out two lessons we have learned from this example: 1. despite the name, the MRO determines the resolution order of the attributes, not only of the methods; 2. the default food for Pythonistas is spam ! (but you already knew that ;-) Having discussed the issue with the local precedence ordering, let me now consider the issue of monotonicity. My goal is to show that both the MRO for classic classes and for new style classes for Python 2.2 are not monotonic. To prove that the MRO for classic classes is non-monotonic is rather trivial, it is enough to look at the diamond diagram: :: C / \ / \ A B \ / \ / D One easily sees the inconsistency:: L(B,P21) = B C # B precedes C : B's methods win L(D,P21) = D A C B C # B follows C : C's methods win! On the other hand, there are no problems with the Python 2.2 and 2.3 MROs, that give both L(D) = D A B C Guido points out in his essay [[#]_] that the classic MRO is not so bad in practice, since one can typically avoids diamonds for classic classes. But all new style classes inherit from object, therefore diamonds are unavoidable and inconsistencies shows up in every multiple inheritance graph. The MRO of Python 2.2 makes break monotonicity difficult, but not impossible. The following example, originally provided by Samuele Pedroni, shows that the MRO of Python 2.2 is non-monotonic:: class A(object): pass class B(object): pass class C(object): pass class D(object): pass class E(object): pass class K1(A,B,C): pass class K2(D,B,E): pass class K3(D,A): pass class Z(K1,K2,K3): pass Here are the linearizations according to the C3 MRO (the reader should verify these linearizations as an exercise and draw the inheritance diagram ;-):: L(A) = A O L(B) = B O L(C) = C O L(D) = D O L(E) = E O L(K1)= K1 A B C O L(K2)= K2 D B E O L(K3)= K3 D A O L(Z) = Z K1 K2 K3 D A B C E O Python 2.2 gives exactly the same linearizations for A, B, C, D, E, K1, K2 and K3, but a different linearization for Z: L(Z,P22) = Z K1 K3 A K2 D B C E O It is clear that this linearization is *wrong*, since A comes before D whereas in the linearization of K3 A comes *after* D. In other words, in K3 methods derived by D override methods derived by A, but in Z, which still is a subclass of K3, methods derived by A override methods derived by D !! This is again a violation of monotonicity which cannot be allowed. On top of that, the Python 2.2 linearization of Z is also inconsistent with local precedence ordering, since the local precedence list of the class Z is [K1, K2, K3] where K2 precedes K3, whereas in the linearization of Z we have that K2 *follows* K3. These problems explain why the 2.2 rule has been dismissed in favor of the C3 rule. Here I give a short Python 2.2 script that allows anybody to compute the 2.3 MRO without risk for his brain. Simply change the last line to play with the various examples I have discussed in this paper. :: # """C3 algorithm by Samuele Pedroni (with readability enhanced by me). Remark: it only works with Python 2.2 (not 2.3+ nor 2.1-).""" class __metaclass__(type): "All classes are metamagically modified to be nicely printed" __repr__ = lambda cls: cls.__name__ class ex_2: "Serious order disagreement" #From Guido class O: pass class X(O): pass class Y(O): pass class A(X,Y): pass class B(Y,X): pass try: class Z(A,B): pass #creates Z(A,B) except TypeError: pass # Z(A,B) cannot be created in Python 2.3+ class ex_5: "My first example" class O: pass class F(O): pass class E(O): pass class D(O): pass class C(D,F): pass class B(D,E): pass class A(B,C): pass class ex_6: "My second example" class O: pass class F(O): pass class E(O): pass class D(O): pass class C(D,F): pass class B(E,D): pass class A(B,C): pass class ex_9: "Difference between Python 2.2 MRO and C3" #From Samuele class O: pass class A(O): pass class B(O): pass class C(O): pass class D(O): pass class E(O): pass class K1(A,B,C): pass class K2(D,B,E): pass class K3(D,A): pass class Z(K1,K2,K3): pass def merge(seqs): print '\n\nCPL[%s]=%s' % (seqs[0][0],seqs), res = []; i=0 while 1: nonemptyseqs=[seq for seq in seqs if seq] if not nonemptyseqs: return res i+=1; print '\n',i,'round: candidates...', for seq in nonemptyseqs: # find merge candidates among seq heads cand = seq[0]; print ' ',cand, nothead=[s for s in nonemptyseqs if cand in s[1:]] if nothead: cand=None #reject candidate else: break if not cand: raise "Inconsistent hierarchy" res.append(cand) for seq in nonemptyseqs: # remove cand if seq[0] == cand: del seq[0] def mro(C): "Compute the class precedence list (mro) according to C3" return merge([[C]]+map(mro,C.__bases__)+[list(C.__bases__)]) def print_mro(C): print '\nMRO[%s]=%s' % (C,mro(C)) print '\nP22 MRO[%s]=%s' % (C,C.mro()) print_mro(ex_9.Z) # Resources --------- .. [#] The thread on python-dev started by Samuele Pedroni: http://mail.python.org/pipermail/python-dev/2002-October/029035.html .. [#] The paper *A Monotonic Superclass Linearization for Dylan*: http://www.webcom.com/haahr/dylan/linearization-oopsla96.html .. [#] Guido van Rossum's essay, *Unifying types and classes in Python 2.2*: http://www.python.org/2.2.2/descrintro.html .. [#] The (in)famous book on metaclasses, *Putting Metaclasses to Work*: Ira R. Forman, Scott Danforth, Addison-Wesley 1999.