In 1970 Royce proposed what is presently referred to as the waterfall model as an initial concept, a model which he argued was flawed (Royce 1970). His paper then explored how the initial model could be developed into an iterative model, with feedback from each phase influencing subsequent phases, similar to many methods used widely and highly regarded by many today. It is only the initial model that received notice; his own criticism of this initial model has been largely ignored. The "waterfall model" quickly came to refer not to Royce's final, iterative design, but rather to his purely sequentially ordered model. This article will use this popular meaning of the phrase waterfall model. For an iterative model similar to Royce's final vision, see the spiral model.
Despite Royce's intentions for the waterfall model to be modified into an iterative model, use of the "waterfall model" as a purely sequential process is still popular, and, for some, the phrase "waterfall model" has since come to refer to any approach to software creation which is seen as inflexible and non-iterative. Those who use the phrase waterfall model pejoratively for non-iterative models that they dislike usually see the waterfall model itself as naive and unsuitable for a "real world" process.
The best-known and oldest process is the waterfall model, where developers (roughly) follow these steps in order:
To follow the waterfall model, one proceeds from one phase to the next in a purely sequential manner. For example, one first completes "requirements specification" — they set in stone the requirements of the software. When the requirements are fully completed, one proceeds to design. The software in question is designed and a "blueprint" is drawn for implementers (coders) to follow — this design should be a plan for implementing the requirements given. When the design is fully completed, an implementation of that design is made by coders. Towards the later stages of this implementation phase, disparate software components produced by different teams are integrated. After the implementation and integration phases are complete, the software product is tested and debugged; any faults introduced in earlier phases are removed here. Then the software product is installed, and later maintained to introduce new functionality and remove bugs.
Thus the waterfall model maintains that one should move to a phase only when its preceding phase is completed and perfected. Phases of development in the waterfall model are thus discrete, and there is no jumping back and forth or overlap between them.
There is a misconception that the process has no provision for correcting errors in early steps (for example, in the requirements). In fact this is where the domain of requirements management comes in which includes change control.
This approach is used in high risk projects, particularly large defense contracts. The problems in waterfall do not arise from "immature engineering practices, particularly in requirements analysis and requirements management." Studies of the failure rate of the specification, which enforced waterfall, have shown that the more closely a project follows its process, specifically in up-front requirements gathering, the more likely the project is to release features that are not used in their current form.
More often too the supposed stages are part of joint review between customer and supplier, the supplier can, in fact, develop at risk and evolve the design but must sell off the design at a key milestone called Critical Design Review (CDR). This shifts engineering burdens from engineers to customers who may have other skills.
The waterfall model however is argued by many to be a bad idea in practice, mainly because of their belief that it is impossible to get one phase of a software product's lifecycle "perfected" before moving on to the next phases and learning from them (or, at least, the belief that this is impossible for any non-trivial program). For example, clients may not be aware of exactly what requirements they want before they see a working prototype and can comment upon it; they may change their requirements constantly, and program designers and implementers may have little control over this. If clients change their requirements after a design is finished, that design must be modified to accommodate the new requirements, invalidating quite a good deal of effort if overly large amounts of time have been invested into "Big Design Up Front". (Thus, methods opposed to the naive waterfall model--such as those used in Agile software development--advocate less reliance on a fixed, static requirements document or design document). Designers may not (or, more likely, cannot) be aware of future implementation difficulties when writing a design for an unimplemented software product. That is, it may become clear in the implementation phase that a particular area of program functionality is extraordinarily difficult to implement. If this is the case, it is better to revise the design than to persist in using a design that was made based on faulty predictions and that does not account for the newly discovered problem areas.
The idea behind the waterfall model may be "measure twice; cut once", and those opposed to the waterfall model argue that this idea tends to fall apart when the problem being measured is constantly changing due to requirement modifications and new realizations about the problem itself. The idea behind those who object to the waterfall model may be "time spent in reconnaissance is seldom wasted".
In summary, the criticisms of a non-iterative development approach (such as the waterfall model) are as follows:
- Many software projects must be open to change due to external factors; the majority of software is written as part of a contract with a client, and clients are notorious for changing their stated requirements. Thus the software project must be adaptable, and spending considerable effort in design and implementation based on the idea that requirements will never change is neither adaptable nor realistic in these cases.
- Unless those who specify requirements and those who design the software system in question are highly competent, it is difficult to know exactly what is needed in each phase of the software process before some time is spent in the phase "following" it. That is, feedback from following phases is needed to complete "preceding" phases satisfactorily. For example, the design phase may need feedback from the implementation phase to identify problem design areas. The counter-argument for the waterfall model is that experienced designers may have worked on similar systems before, and so may be able to accurately predict problem areas without time spent prototyping and implementing.
- Constant testing from the design, implementation and verification phases is required to validate the phases preceding them. Constant "prototype design" work is needed to ensure that requirements are non-contradictory and possible to fulfill; constant implementation is needed to find problem areas and inform the design process; constant integration and verification of the implemented code is necessary to ensure that implementation remains on track. The counter-argument for the waterfall model here is that constant implementation and testing to validate the design and requirements is only needed if the introduction of bugs is likely to be a problem. Users of the waterfall model may argue that if designers (et cetera) follow a disciplined process and do not make mistakes that there is no need for constant work in subsequent phases to validate the preceding phases.
- Frequent incremental builds (following the "release early, release often" philosophy) are often needed to build confidence for a software production team and their client.
- It is difficult to estimate time and cost for each phase of the development process without doing some "recon" work in that phase, unless those estimating time and cost are highly experienced with the type of software product in question.
- The waterfall model brings no formal means of exercising management control over a project and planning control and risk management are not covered within the model itself.
- Only a certain number of team members will be qualified for each phase; thus to have "code monkeys" who are only useful for implementation work do nothing while designers "perfect" the design is a waste of resources. A counter-argument to this is that "multiskilled" software engineers should be hired over "specialized" staff.
