Bertrand Meyer is a software pioneer whose activities have spanned both the academic and business worlds. He is currently the Chair of Software Engineering at ETH, the Swiss Institute of Technology. He is the author of numerous papers and many books, including the classic Object-Oriented Software Construction (Prentice Hall, 1994, 2000). In 1985, he founded Interactive Software Engineering, Inc., now called Eiffel Software, Inc., a company which offers Eiffel-based software tools, training, and consulting.
On September 28, 2003, Bill Venners conducted a phone interview with Bertrand Meyer. In this interview, which is being published in multiple installments on Artima.com, Meyer gives insights into many software-related topics, including quality, complexity, design by contract, and test-driven development.
Bill Venners: The aim of test-driven development is to help programmers avoid bugs, to get systems that are robust. But one aspect of a system that is not just robust, but also reliable, is that when things do go wrong, either because of a bug or because of circumstances outside of the system, the system can deal with the problem without requiring an administrator or user to solve the problem. How can we create systems that deal with failure autonomously so that humans don't have to step in? We have the tool of exceptions, but what do we do with them? And what do we do when a contract assertion is false?
Bertrand Meyer: The really deep and final answer is it depends. There are really two approaches. One approach is to say, this problem simply shouldn't happen. If it ever does happen, the best you can do is shut your system down, fix the bug, and restart it. Some people take this approach, but it is probably not sustainable for telephone system. If you're AT&T and you're handling millions of telephone calls in your system and suddenly an invariant is violated, you're not going to shut off the AT&T network. For other kinds of systems, however, it is probably the most reasonable thing to do. Some problem has not been caught in debugging. It should have never happened, so they just stop the whole system, correct the defect, and restart. That is one approach. It's rather extreme. The other approach is to do essentially fault-tolerant computing. Not hardware fault tolerance, but software fault tolerance, which is a relatively new option. The term has been around a long time, but the approach hasn't been practiced that much.
The first thing to do if you have a problem is obviously to log it. Typically when you have an assertion violation during operation, not during debugging, but during operation, it's an indication that something needs to be fixed in the software. The defect cannot be left to stand in the long term. What you do in the short term is try to recover as reasonably as you can. And indeed, that's where exception handling comes in. The approach I suggest for exception handling is more low profile than the approach that seems to have become popular these days. In most recent programming languages, exceptions are a normal part of life. For example, exceptions figure prominently in the specification of operations. The exception handling strategy that I've pushed for is more low profile in the sense that it views exceptions really as what happens when everything else has failed, and you don't have much of a clue as to what is going on except that something is seriously wrong. The best you can really do is to try to restart in a clean state.
The exception mechanism in Eiffel is quite different from those that exist in other languages, and it surprises many people. When you have an exception in Eiffel, you only have two ways of reacting to it. The reason that you only have two ways is that exceptions are something that you don't want to see happening. So it's quite different from the approach that says exceptions are special cases that we are going to expect and process in a special way. In Eiffel, exceptions are really a sign that something quite wrong has happened, so you have only these two ways of reacting. One is to accept that you can not do anything better, and to just pass on the problem up to the next routine up the call chain, which of course will be faced with the same dilemma. That is often, especially for operations deep down in the call chain, the only real world reaction, because the operation does not have enough context to do anything smarter. The other reaction is, if you actually do have some elements that enable you to think you're smarter, to attempt to fix the condition that led to the exception and try the same operation again after correcting the context. This is the kind of mechanism that provides direct support for what I was calling software fault tolerance, and I think it can be quite effective provided it's used with reason, that is to say, not as another algorithmic mechanism, but as a mechanism of last resort.
Bill Venners: One of the recommendations you make in your book Object- Oriented Software Construction that has always surprised me is the notion that under no circumstances shall the body of a routine check for the precondition of the routine. I understand that you don't want to have the same condition being tested in multiple places, both the client and the supplier, as you called it, but you also said, "If the client's part of the contract is not fulfilled, that is to say if the class does not satisfy precondition, then the class is not bound by the post condition." So the supplier class can either throw an exception or return a wrong result. I tend to prefer defining exceptions that will be thrown if preconditions are broken, because then the behavior of the routine is specified even under broken preconditions. The behavior of the routine is fully specified. That's my intuitive preference. Could you describe your reasoning for recommending the other approach?
Bertrand Meyer: I don't think you're expressing a preference with respect to language rules, but rather, a preference for a design style that puts more responsibility on the supplier. That is to say, you prefer a design style in which the supplier in effect extends the specification to account for more cases. In the end, it's really a matter of design style. Personally, I have found it far more effective to define for each operation the conditions under which it will work, and leave the behavior undefined in other cases. I think in the end it's a more effective approach.
If you take the viewpoint of the client, what happens typically is this: You are writing some software and you instantiate a class. Assuming you are programming in a context in which the only way to use the class is through it's specification, which includes the contract, you'll see the pre-condition. If you are a reasonable client programmer, if you know what you are doing, you'll know to program against the precondition. Of course, you may not like the precondition. You may find it's too demanding, but that's life. It's part of the contract. I think that style works pretty well.
The other style is a bit more problematic. For example, I recently looked at some collections classes from .NET. You can see the interface for a certain library class, but you don't necessarily see the exceptions. If you're careful and serious, you will go deeper into the documentation of the class and see that some exceptions can be raised in certain cases. If you want to process such cases, you'll have to write a try catch structure in your own client code. That structure is actually more complicated than dealing with the negation of the precondition in the first place, because you're trying to do some task, and after the fact you have to catch possible violations and try to recover. So you will actually end up doing the same thing either way, but it's more complicated to deal with it after the fact than if you just checked the precondition in the first place. Likely, however, many programmers are simply not going to take the trouble to look at these exception clauses, or they might miss some of them. If something wrong happens, some exception will occur, and there's no guarantee at all whether and how the exception is going to be processed. Is it going to be processed properly? Is it going to be passed back up the call chain to a routine that has no clue about what actually happened? So it seems to me that this approach is really a way to ignore the problem, to put your head in the sand and pretend that nothing is going to happen, when in fact things can happen. It is better to tackle the problem head on, to have a specification that explicitly says what is expected of the caller to get correct treatment, and to make sure before the call that these conditions are satisfied.
Bill Venners: My last question is about reuse. You have talked a lot about reuse, and seem to value reuse. I find that I reuse classes all the time in standard APIs, but that in the context of everyday programming tasks, it's often a better approach in my experience to solve the specific problem. In their everyday programming tasks, how much should programmers worry about reuse of the components or pieces of software they are building?
Bertrand Meyer: That's a very good point. I think it's really not a problem for the programmers. It's really a problem for the businesses, for the corporations. That is to say, if you are a programmer working in the usual conditions of typically a lot of stress, pressure, and deadlines, then it can actually be detrimental to focus too much on making your software reusable. You may even be accused of not doing your work properly, and that accusation might be in part at least justified. Your job is to produce is a set of products by a certain time at a certain price with certain functionality. That's what you have to do, and reuse can wait. Reuse is usually not part of the specification.
Of course, you should be a good programmer. You should do things carefully. Contrary to what I think the XP people are saying—I may be misinterpreting here, but I think it is a major point of disagreement I have with XP—if you have a way to do something more generally and another way to do it more specifically, at roughly equal cost, you should always try to do the more general solution. So you should always think about reuse, but really it would be improper in many cases to make something reusable. You would not be fulfilling your duty to the group and to the corporation if you were spending too much time on generalizing for future benefit as opposed to the immediate benefit of the project.
It's really in the end a question for the corporation. Does the corporation want to spend a little more money and time on generalization once the product has been delivered or the milestone achieved? I think I understood only recently the difference between this and the idea of refactoring. A few years ago I published a book called Object Success, a presentation of object technology for management, in which I talked a lot about reuse. In particular, I pushed this idea that the software lifecycle should allow for an explicit step of generalization. The idea I think is very simple: put management in front of its responsibilities. Many companies that will say, we don't have the time to do this. We just want to deliver a product, and we don't have time for any extra effort to generalize the software. It's not part of our charter. And I would say, that's fine. It has the advantage of being completely open and frank and conscious, as opposed to the unconscious decisions that are far too often made in software environments.
On the other hand, I think that a more progressive and forward looking software environment would say, yes, we are going to devote a small but reasonable part of our project, five to ten percent, to make sure that after the project is delivered we not only clean up the code but perform some generalization to prepare for the next project, to prepare for reuse. I think in the end that will be more effective than constant refactoring. First you meet your deadlines, and then you have a period of your work that is officially devoted to making things better. Then you do the refactoring, but with a specific focus on reuse. That refactoring is an official part of your job description, not something that you do on the side, almost hiding it from your manager, pretending that you're working on some other project In the end I think it's a management decision. Are you are you not willing to spend more in order to not only deliver to your customers and users what they expect now, but also to provide for better software development in the future. It is fundamentally in a business decision.
Come back Monday, March 15 for the first installment of a conversation with Luke Hohman. If you'd like to receive a brief weekly email announcing new articles at Artima.com, please subscribe to the Artima Newsletter.
Bertrand Meyer is the author of Object-Oriented Software Construction, which is available on Amazon.com at:
Bertrand Meyer's Home Page at ETH, the Swiss Institute of Technology:
The Grand Challenge of Trusted Components, presented by Bertrand Meyer at the International Conference on Software Engineering, May 2003:
The Structure and Interpretation of Computer Programs, by Harold Abelson and Gerald Jay Sussman with Julie Sussman:
Find out more about Eiffel at:
The Eiffel language FAQ:
The 2001 Interview with Programming Expert Bertrand Meyer in InformIT:
(Gratuitously long URL omitted...)
Design by Contract by Example, by Richard Mitchell and Jim McKim:
Object Success: A Manager's Guide to Object-Oriented Technology And Its Impact On the Corporation, by Bertrand Meyer:
Bill Venners is president of Artima Software, Inc. and editor-in-chief of Artima.com. He is author of the book, Inside the Java Virtual Machine, a programmer-oriented survey of the Java platform's architecture and internals. His popular columns in JavaWorld magazine covered Java internals, object-oriented design, and Jini. Bill has been active in the Jini Community since its inception. He led the Jini Community's ServiceUI project that produced the ServiceUI API. The ServiceUI became the de facto standard way to associate user interfaces to Jini services, and was the first Jini community standard approved via the Jini Decision Process. Bill also serves as an elected member of the Jini Community's initial Technical Oversight Committee (TOC), and in this role helped to define the governance process for the community. He currently devotes most of his energy to building Artima.com into an ever more useful resource for developers.