Design Patterns Dennis Mancl Alcatel-Lucent – Bell Labs November 28, 2007
Transcription
Design Patterns Dennis Mancl Alcatel-Lucent – Bell Labs November 28, 2007
Design Patterns Dennis Mancl Alcatel-Lucent – Bell Labs November 28, 2007 1 1 Design Patterns Outline • What is a pattern? – The Design Patterns Book – One example pattern: Singleton – Patterns versus Idioms • • • • • • Wrapping with Facade objects Remote access with Proxy Observer Analysis patterns Architectural patterns Enterprise Architecture patterns 2 2 What is a pattern? • A pattern is a solution to a problem in a context • The problem and context come from some domain (software development, project management, …) – The problem and context contain some “unresolved forces” that need to be addressed. • The solution is a proposed way to resolve the problem Name: Information Expert Problem: How to decide which class or object to assign responsibilities to? Context: A responsibility usually requires some information or data for its fulfillment – information about other objects, an object’s own state, the world around an object, and so on. Solution: Assign a responsibility to the class that has the information needed to fulfill it. (from Applying UML and Patterns, third edition, Craig Larman) 3 Also see the description of “Pattern Language” in the Wikipedia -http://en.wikipedia.org/wiki/Pattern_language. 3 What is a pattern? • Patterns are used to capture knowledge and experience – for example, what to think about when making design choices • Not a cookbook • Each pattern might trigger others Name: Loose Interfaces Problem: To avoid development bottlenecks, we need to be able to limit the effect that one team’s work will have on another. Context: Interfaces need to be somewhat flexible if the teams are working quickly. Requirements may be changing rapidly, and it may be difficult to communicate between geographically distributed subteams. Solution: Limit the number of explicit, static interfaces. Define larger-grained interfaces that allow developers to code against interfaces defined early, but that do not overly constrain functionality. Consider using the Hierarchy of Factories pattern. (from Organizational Patterns of Agile Software Development, Coplien and Harrison) 4 4 Patterns in non-software domains • Name: ChocolateChipRatio • Context: You are baking chocolate chip cookies in small batches for family and friends • Consider these patterns first: SugarRatio, FlourRatio, EggRatio • Problem: Determine the optimum ratio of chocolate chips to cookie dough • Solution: Observe that most people consider chocolate to be the best part of the chocolate chip cookie. Also observe that too much chocolate may prevent the cookie from holding together, decreasing its appeal. Since you are cooking in small batches, cost is not a consideration. Therefore, use the maximum amount of chocolate chips that results in a sturdy cookie. • Consider next: WalnutRatio or CookingTime or FreezingMethod (from Wikipedia: http://en.wikipedia.org/wiki/Pattern_language) 5 This example shows that there are multiple patterns in a Cookie pattern language, but this pattern focuses on one of the critical characteristics of good chocolate chip cookies – getting the amount of chocolate right. Personally, I always follow the Nestle recipe, which is the recipe on the bag of Nestle’s toll house morsels – their ratio is just about right to satisfy the competing forces of good taste and solid cookie architecture. 5 Software patterns • The first “Object Oriented Design Patterns” are found in the book – Design Patterns by Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides (known as “Gang of 4” or “GOF”) – 23 patterns: three categories (Creational, Structural, Behavioral) – Problems: • control the construction of objects • create organized collections of objects • divide responsibilities among objects for better encapsulation and easier modification – Examples in C++ and Smalltalk – The patterns also apply to other languages (Java, Visual Basic, Perl) 6 The Design Patterns book first appeared in fall 1994. The book was a big hit at the OOPSLA conference (sponsored by ACM SIGPLAN) – the Addison-Wesley booth was completely sold out of copies of the book. Many of the other books on design patterns make reference to these 23 patterns – either by repeating them, extending them, or discussing some specializations of the patterns for certain contexts. 6 Singleton Pattern • Name: Singleton • Problem: We need to restrict the creation of new objects for a particular class: only one instance. • Context: Other objects in the design need a single point of access to use the object. • Solution: Define a “static member function” within the class that returns a pointer (or reference) to the Singleton object. The static member function will create the object the first time it is called. All constructors are made private to prevent other objects from being created. 7 Singleton is the simplest pattern in the book. It is relatively easy to implement in C++. The access to the object is controlled: by putting the “instance” in the private section of the class, any function outside of the class can’t directly access the object – the access must be done through the advertised interface (the getInstance() static member function). Note: The actual “construction” of the singleton object is deferred until the first time the application tries to access it – see the next page. 7 More about Singleton • Simple version of Singleton (in C++, without multi-thread support): MySingletonClass – instance : MySingletonClass * other attributes … + getInstance() : MySingletonClass * – MySingletonClass() other operations … underline means static data member if (instance == 0) { instance = new MySingletonClass(); } return instance; • Access the Singleton like this: MySingletonClass::instance() • Different implementations of the Singleton pattern for Java, multithreaded support, allocation in special memory, and so on. 8 Suppose you need a Singleton in a multi-threaded context. In that case, the instance == 0 test plus the call to the new operator in the getInstance() function is a “critical section”. If two threads try to do this operation at the same time, the results are unpredictable. You *might* wind up with two instances of the singleton object. A solution to this problem can be found in Design Patterns Explained by Alan Shalloway and James R. Trott. if (instance == 0) { lock.acquire(); if (instance == 0) { instance = new MySingletonClass(); } lock.release(); } return instance; 8 Patterns and Idioms • What have we done? We have used a “design trick” that solves a specific problem • Some standard design tricks are called “idioms” – they tend to be small, tricky, and language specific – example: virtual destructors in C++ – example: anonymous inner classes in Java • How are patterns different from idioms? – Patterns are more about design – Singleton is useful in Java, Smalltalk, Visual Basic, Perl, … – different “realizations” of patterns in different languages 9 There is some overlap between idioms and design patterns. But for the most part, idioms are aimed at more concrete circumstances, and they tend to be written as “rules”: • define the copy constructor, assignment operator, and destructor for C++ classes that manage dynamic resources (for example, classes that allocate dynamic memory regions, open file descriptors, or open databases) • always define a virtual destructor for C++ classes that contain other virtual functions • never return a reference to an object on the stack in a C++ function 9 Patterns in the Design Patterns book • 23 Design Patterns: three categories Creational Structural Abstract Factory Adapter Chain of Responsibility Observer Builder Bridge Command State Interpreter Strategy Factory Method Composite Behavioral Prototype Decorator Iterator Template Method Singleton Facade Mediator Visitor Flyweight Memento Proxy 10 10 Sources of information • So, you want to learn how to use these design patterns – to make your OO designs better • Where to look to learn more: – the book – A Learning Guide to Design Patterns (http://www.industriallogic.com/papers/learning.html) – Pattern Stories Wiki (http://wiki.cs.uiuc.edu/PatternStories) – http://c2.com/cgi/wiki?DesignPatterns – other books: • • • • Pattern-Oriented Software Architecture by Frank Buschmann et. al. Design Patterns Java Workbook by Steven John Metsker Refactoring to Patterns by Joshua Kerievsky Design Patterns Explained by Alan Shalloway and James R. Trott 11 New books, articles, and courses on patterns are being written every day. If you are an ACM member (ACM is Association for Computing Machinery), you can take free online training courses, including a introduction to Design Patterns: http://pd.acm.org/course_slc.cfm?pre=ITDT&trk=654&crs=WT-1230-180 11 Why learn the Design Patterns? • Your own designs will improve – borrowing well-tested ideas – pattern descriptions contain some analysis of tradeoffs • You will be able to describe complex design ideas to others – assuming that they also know the same patterns • You can use patterns to “refactor” existing code – refactoring is improving the structure of existing code without adding new functionality – some techniques for transforming and adapting legacy code are based on design patterns 12 12 Facade Pattern • Problem: The application needs a simple interface to a complex subsystem – the subsystem might have been written by someone else – you don’t want the entire development team to go back to the subsystem documentation with lots of questions • Context: – it is important to control the dependencies between the application and the complex subsystem – you want to reduce the effort to maintain the system • Solution: Define a single Facade class – the Facade class has knowledge of the internal details of the subsystem, but the Facade class provides a simple to use interface for the application – each Facade public function might call many subsystem operations 13 13 Facade diagram • A Facade class wraps a bunch of operations on other classes (or a bunch of legacy code operations) into a convenient package application application calls some of the Facade class operations Facade class scanner pointers to other data objects within a subsystem get_val1() build_trans2() do_trans3() commit_trans4() the Facade accesses some of the internals of a complex subsystem to implement its operations database interface parser formatter 14 A Facade class has some interesting internal structure: • the data section of a Facade class is usually more than just a simple pointer • the implementation of the Facade member functions might need to interact with many classes and/or functions in the subsystem that it is hiding The Facade pattern is the most commonly used pattern for burying a non-OO legacy subsystem. 14 Planning a Facade • Part of the OO design process: create a simplified model of the classes in the overall system – using CRC cards (informal technique using index cards, one card per class) – using UML Class Diagrams • Look for a group of classes that collaborate closely together: call them a “subsystem” • Try to “put the subsystem behind a wall” – define a single interface class that defines an API for the subsystem 15 15 Proxy example • A Proxy object is a “stand-in” for an object that might be: – located on another machine (example: CORBA objects, Java RMI) – or it might be large and complicated, so you want to defer building it – or it might require a special access method (to check the users’ permissions) CORBA example client application Stub objects have full public interface, but no data. They just push a message over to the server. ORB (server) CORBA IDL stubs IDL skeletons Interface Repository Object Implementations network requests responses Generated by a tool Developers fill in details 16 In the Proxy pattern, the Proxy object doesn’t necessarily have a simple way to access the information in the class that stands behind the Proxy object: it might involve a complex communication protocol instead of just following a pointer. For example, CORBA implementations usually generate client-side “stub” classes (which are simple Proxy classes) directly from an IDL specification of a class). The operations of these stub classes will send messages to the “real” object implementations that live on a server machine. The Object Request Broker (ORB) on the client machine knows how to parse the messages from stub classes. The ORB will pick out the right operations from the CORBAgenerated code (the IDL skeletons and Interface Repository) and the user-constructed code (the Object Implementations) to return the right information to the client. Examples of non-distributed-object related Proxy objects include: • bitmap objects (creating or downloading a bitmap might be deferred until the application knows what section of the bitmap is needed) • sparse matrix objects (you might not compute all elements of an array immediately) 16 Proxy pattern • Problem: The application needs to operate on a distant object • Context: Need a placeholder for an object – if the actual object is far away (on another computer or on a disk), the placeholder should be in the local address space • Solution: Define a Proxy class: this class will have the same public interface as the real class, and when a Proxy operation is called, the internals of a Proxy object will arrange for the same function to be called on a real object – Proxy might send interprocess or intermachine messages – Proxy might resurrect an object that has been written to disk – Proxy might acquire special permissions to access a protected real object 17 17 How to tell the difference? • How can I tell which one I am using? Facade or Proxy? – Both patterns are separating the implementation details from an abstract interface – But the intent is different for each pattern • Facade: the application wants to use services of an entire subsystem or package (usually tens or hundreds of objects in the subsystem) • Facade is also a common way to interface with non-OO modules • Proxy provides convenient access to an object – the object might be on disk, in another process, on another computer 18 18 Observer – a useful design technique • Observer pattern always has at least 2 classes: Subject and Observer – TemperatureSensor – maintains information about the current temperature – TempReportWebPage – displays temperature value on a web page • Each temperature value might appear on multiple pages • Update each page when the TemperatureSensor changes state page1: TempReportWebPage temp: TemperatureSensor report_URL : string Update page2: TempReportWebPage call Update() on each page report_URL : string Update page3: TempReportWebPage report_URL : string Update 19 Observer is useful because there is often a separation between state and behavior in certain designs. If an object’s state is changed, there may be several objects that expect to be informed of that state change. Note that in this case, the “action to be taken” on a state change of TemperatureSensor usually belongs to the “observer” class (TempReportWebPage). The TemperatureSensor class doesn’t know anything about web pages – TemperatureSensor only keeps a list of other objects that depend on the state of the sensor. (Note: In the example on this slide, the TemperatureSensor object is responsible for updating multiple web pages. This is a common occurrence – you might have a system that records temperatures from various reporting stations, but the screens to “view” the data might be interested in different subsets of the reporting stations… all of the airports in a particular state, all airports serving a metropolitan area, all airports that have more than 500 flights per day, and so on.) 19 Simple class diagram for Observer • In the Observer Pattern: there are links between two kinds of concrete classes – subjects and observers • • Each observer object (such as TempReportWebPage) needs to register by calling the Attach() operation on a subject object (like TemperatureSensor) Each observer object will have its Update() operation called whenever its subject changes state ConcreteSubject observers ConcreteObserver subjectState 1 0..* observerState GetState() Update() ModifyState() Attach(Observer) observerState = Detach(Observer) subject->GetState(); Notify() for all o in observers { o ->Update() } subjectState = newState; Notify(); 20 This is a simplified version of the “general picture” of the Observer pattern. This pattern can be made more general by adding abstract classes… see the next page. 20 Complete class diagram for Observer • • • Observer is an abstract interface that each ConcreteObserver must implement (must implement an Update() function) Observer objects still register by calling the Attach() operation on a ConcreteSubject object Each ConcreteObserver object will have its Update() operation called whenever its ConcreteSubject changes state Subject Attach(Observer) Detach(Observer) Notify() observers Observer 1 0..* Update() for all o in observers { o ->Update() } ConcreteSubject subjectState GetState() return subjectState ConcreteObserver observerState Update() observerState = subject->GetState(); 21 This is the “general picture” of the Observer pattern (from the Design Patterns book). It has been generalized: the ConcreteSubject gets the “notification” functionality from its base class Subject. Within the implementation of the ConcreteSubject class, there will be one or more calls to the Subject::Notify() function – whenever the Subject determines that the Observers need to be notified. Observer is an “abstract class” – it defines the interface that all ConcreteObservers must implement. 21 Why update the Observer objects? • • In the Observer Pattern, there are some objects that “care about” the state changes within other objects The ConcreteObserver::Update() function might do these things: – repaint a user interface – check for exceeding thresholds • So the ConcreteObserver is a place to put operations that might clutter up a domain class – example: any “View” class • Or the ConcreteObserver is a place to put a command handler for a user interface class – button and menu handlers can be implemented as Observers of GUI objects 22 22 Motivation for Observer • An early OO user interface architecture: Model-ViewController – A model, is an object representing data (often data about an object in the real world) – A view is some form of visualization of the state of the model. – A controller offers facilities for the user to change the state of the model (clicking on buttons, sliding a slider, typing in a text box, etc.) • Some elements within the View observe the Model – if the Model changes state, then the View may need to update its objects and pass the changes to the user interface implementation • The Model observes the Controller – after the Controller accepts command input from the user, it needs to notify the Model to update its state View A 5 Z Model event proc1 event proc2 Controller 23 Model-View-Controller (MVC) is a classic “separation of concerns” within a design. It is possible to implement this with a single Model class, a single View class, and a single Controller class, or with each part of the architecture implemented using a number of related classes. A good concise explanation of Model-View-Controller can be found on the OO Tips website: http://ootips.org/mvc-pattern.html. See also http://c2.com/cgi/wiki?WhatsaControllerAnyway. Note that there are some frameworks that implement a variation of MVC: Document-View. This combines the Controller and the View into a single Document class. One popular implementation of the Document-View architecture is the Microsoft Foundation Classes (MFC). 23 Observer implementation • A Subject class will usually keep a linked list of pointers to its registered Observers – linked list: multiple View objects might observe the same Model, and the linked list makes it easier when one of the middle objects wants to Detach (to stop receiving notifications) – each pointer in the list is of type Observer* (abstract type defining the Update() operation), because the concrete types for different observers is often quite different • Is this necessary? – No. In some cases, you might know that all observers have the same type. Then the list might declared to contain ConcreteObserver* pointers. – In fact, the linked list could be replaced by a fixed sized pointer table or a dynamic sized array (std::vector in C++) – especially if the set of observers is relatively static 24 24 Analysis patterns • Analysis patterns are a set of patterns that are used in doing the initial problem analysis: – They help answer the question “what objects should I define in my system?” • The Quantity pattern is from the book Analysis Patterns by Martin Fowler – Recording measurements and manipulating results might be error-prone – Each value really should be recorded with its units: • A Money object will have both a number and an identifier to say which currency: [19.95, “US Dollars”]; [700, “Euros”]; [100, “Yuan”] • Length and weight also need units: [100, “miles”]; [15.5, “kg”] 25 Analysis Patterns are just like Design Patterns – they are simple solutions to recurring problems. Each Analysis Pattern has problem and a context, and it identifies certain common solutions. 25 Justification for the Quantity pattern • A frequent problem – someone tries to perform an invalid operation on two different types of quantities: – adding apples to oranges, people to money, dates to time intervals – conversion mistakes: adding dollars to euros, inches to feet – performing an average of a mixed bag of objects (this should never be legal) • Using explicit units in the design makes it easier for someone else to understand the software later – what does this number mean?? 26 26 Architecture patterns • You can save a lot of design work by reusing some welltested architectures: – Patterns of Software Architecture: the first book on “architecture patterns” • An example: Client-Dispatcher-Server – Problem: You are designing a “video on demand” service – Your customers can order different movies, and you have video servers in the central office that can play the movies – When requests come in, the system needs to decide which server will handle each request – There is a standard “division of responsibilities” in the design 27 27 Client-Dispatcher-Server example • initial stimulus send_request(s) A Customer wants a particular Service, so Customer::send_request() is c: Customer 1: receive_request(c) • called. This will invoke the Service (it calls the Service::receive_request() s: Service 1.1: sn = obtain_channel(c) • :ServiceNode Assigner 1.1.2: add NodeUser(c) 1.1.1*: get Capacity_and_Load() • function for a specific Service object). The Service object wants either to create a new ServiceNode or (more likely) to assign an existing ServiceNode to the Customer. The ServiceNodeAssigner class makes the decision about which ServiceNode to assign and configure. sn: ServiceNode 1.1.4: configure Service() 28 28 Client-Dispatcher-Server example • This design uses the “Client-Dispatcher-Server” pattern: – three classes share responsibility for managing communication • the Server provides a set of services, but it initially registers itself with the Dispatcher • the Client will invoke the services provided by one or more Servers, but it will contact the Dispatcher each time it needs a service (to help it find the right Server) • the Dispatcher tracks the Servers and responds to the Clients that are trying to locate a service • The design is relatively clean: – Clients have a single logical point of contact – the Servers don’t need to know anything about the assignment policy 29 29 Enterprise Architecture Patterns • Martin Fowler’s excellent book: Patterns of Enterprise Application Architecture – Enterprise Applications – software that uses databases (concurrent access to persistent data) – How do you write a Java application to manipulate and display data from a database? – How can you make it easier to modify the application to meet new needs without breaking the existing code? • Some answers: – use layers, build facades, define proxies, encapsulate data records, and encapsulate transactions 30 30 Example enterprise patterns • How do I control the formatting of my Web pages? – If I want to edit the page and put hooks in for dynamic data? – use Template View (HTML page with markers for varying information) – The page is a transform of domain data? – use Transform View – If I need to make general “look-and-feel” changes across my web site? – use Two Step View – Multiple appearances for the same logical screen? – use Two Step View • How do I interact with the database? – If I have a domain model that corresponds to the database tables? – use Active Record (encapsulating database access and domain logic into one class) – If I have a rich domain model? – use Data Mapper – If I’m using a Table Module? – use Table Data Gateway 31 These examples are taken from the “Cheat Sheet” in the back cover of the Martin Fowler Patterns of Enterprise Application Architecture book. 31 Summary • Design Patterns: an important set of object oriented design concepts – these patterns are useful in many applications – every pattern has a documented set of “Consequences” • Some useful sources on design patterns: – – – – http://hillside.net/patterns/onlinepatterncatalog.htm http://www.martinfowler.com/articles/enterprisePatterns.html http://java.sun.com/blueprints/corej2eepatterns http://www.headfirstlabs.com/books/hfdp/ 32 32