Google首席工程师Joshua Bloch谈如何设计优秀的API

Google首席工程师Joshua Bloch谈如何设计优秀的API

How to Design a Good API and Why it Matters


Why is API Design Important?

  • APIs can be among a company's greatest assets
_ Customers invest heavily: buying, writing, learning
_ Cost to stop using an API can be prohibitive
_ Successful public APIs capture customers

  • Can also be among company's greatest liabilities
_Bad APIs result in unending stream of support calls
  • Public APIs are forever - one chance to get it right

Why is API Design Importantto You?

  • If you program, you are an API designer
_Good code is modular–each module has an API
  • Useful modules tend to get reused
_ Once module has users, can’t change API at will
_ Good reusable modules are corporate assets
  • Thinking in terms of APIs improves code quality

Characteristics of a Good API

  • Easy to learn
  • Easy to use, even without documentation
  • Hard to misuse
  • Easy to read and maintain code that uses it
  • Sufficiently powerful to satisfy requirements
  • Easy to extend
  • Appropriate to audience


I. The Process of API Design
II. General Principles
III. Class Design
IV. Method Design
V. Exception Design
VI. Refactoring API Designs

The Process of API Design

Gather Requirements–with a Healthy Degree of Skepticism

  • Often you'll get proposed solutions instead
_Better solutions may exist
  • Your job is to extract true requirements
_Should take the form ofuse-cases
  • Can be easier and more rewarding to build something more general

Start with Short Spec–1 Page is Ideal

At this stage, agility trumps completeness
Bounce spec off as many people as possible
_Listen to their input and take it seriously
If you keep the spec short, it’s easy to modify
Flesh it out as you gain confidence
_This necessarily involves coding

Write to Your API Early and Often

Start before you've implemented the API
_ Saves you doing implementation you'll throw away
Start before you've even specified it properly
_ Saves you from writing specs you'll throw away
Continue writing to API as you flesh it out
_ Prevents nasty surprises
_ Code lives on as examples, unit tests

Writing to SPI is Even More Important

Service Provider Interface (SPI)
_Plug-in interface enabling multiple implementations
_Example: Java Cryptography Extension (JCE)
Write multiple plug-ins before release
_ If you write one, it probably won't support another
If you write two, it will support more with difficulty
_ If you write three, it will work fine
Will Tracz calls this “The Rule of Threes”
(Confessions of a Used Program Salesman, Addison-Wesley, 1995)

Maintain Realistic Expectations

Most API designs are over-constrained
_ You won't be able to please everyone
_Aim to displease everyone equally
Expect to make mistakes
_ A few years of real-world use will flush them out
_ Expect to evolve API

General Principles

API Should Do One Thing and Do it Well

Functionality should be easy to explain
_ If it's hard to name, that's generally a bad sign
_Good names drive development
_Be amenable to splitting and merging modules

API Should Be As Small As Possible But No Smaller

API should satisfy its requirements
When in doubt leave it out
_ Functionality, classes, methods, parameters, etc.
_ You can always add, but you can never remove
Conceptual weight more important than bulk
Look for a good power-to-weight ratio

Implementation Should Not Impact API

Implementation details
_ Confuse users
_Inhibit freedom to change implementation
Be aware of what is an implementation detail
_Do not overspecify the behavior of methods
_ For example: do not specify hash functions
_All tuning parameters are suspect
Don't let implementation details “leak” into API
_ On-disk and on-the-wire formats, exceptions

Minimize Accessibility of Everything

Make classes and members as private as possible
Public classes should have no public fields(with the exception of constants)
This maximizes information hiding
Allows modules to be used, understood, built,tested, and debugged independently

Names Matter–API is a Little Language

Names Should Be Largely Self-Explanatory
_ Avoid cryptic abbreviations
Be consistent–same word means same thing
_ Throughout API, (Across APIs on the platform)
Be regular–strive for symmetry
Code should read like prose
if (car.speed() > 2 * SPEED_LIMIT)
generateAlert("Watch out for cops!");

Documentation Matters

Reuse is something that is far easier to say than
to do. Doing it requires both good design and
very good documentation. Even when we see
good design, which is still infrequently, we won't
see the components reused without good documentation.
 - D. L. Parnas, _Software Aging. Proceedings of 16th International Conference Software Engineering, 1994

Document Religiously

Document every class, interface, method,constructor, parameter, and exception
_ Class: what an instance represents
_Method: contract between method and its client
_ Preconditions, postconditions, side-effects
_ Parameter: indicate units, form, ownership
Document state space very carefully

Consider Performance Consequences of API Design Decisions

Bad decisions can limit performance
_ Making type mutable
_ Providing constructor instead of static factory
_ Using implementation type instead of interface
Do not warp API to gain performance
_ Underlying performance issue will get fixed,but headaches will be with you forever
_ Good design usually coincides with good performance

Effects of API Design Decisions on Performance are Real and Permanent

Component.getSize()returns Dimension
Dimension is mutable
EachgetSize call must allocateDimension
Causes millions of needless object allocations
Alternative added in 1.2; old client code still slow

API Must Coexist Peacefully with Platform

Do what is customary
_Obey standard naming conventions
_Avoid obsolete parameter and return types
_Mimic patterns in core APIs and language
Take advantage of API-friendly features
_ Generics, varargs, enums, default arguments
Know and avoid API traps and pitfalls
_Finalizers, public static final arrays

Class Design

Minimize Mutability

Classes should be immutable unless there’s a good reason to do otherwise
_Advantages: simple, thread-safe, reusable
_Disadvantage: separate object for each value
If mutable, keep state-space small, well-defined
_Make clear when it's legal to call which method

Bad:Date, Calendar
Good: TimerTask

Subclass Only Where It Makes Sense

Subclassing implies substitutability (Liskov)
_Subclass only when is-a relationship exists
_Otherwise, use composition
Public classes should not subclass other public classes for ease of implementation
Bad: Properties extends Hashtable
Stack extends Vector
Good: Set extends Collection

Design and Document for Inheritance or Else Prohibit it

Inheritance violates encapsulation (Snyder, ‘86)
_Subclass sensitive to implementation details of superclass
If you allow subclassing, documentself-use
_How do methods use one another?
Conservative policy: all concrete classes final

Bad: Many concrete classes in J2SE libraries
Good: AbstractSet, AbstractMap

Method Design

Don't Make the Client Do Anything the Module Could Do

Reduce need for boilerplate code
_ Generally done via cut-and-paste
_ Ugly, annoying, and error-prone
import org.w3c.dom.*;
import javax.xml.transform.*;
import javax.xml.transform.dom.*;
// DOM code to write an XML document to a specified output stream.
private static final void writeDoc(Document doc, OutputStream out)throws IOException{
try {
Transformer t = TransformerFactory.newInstance().newTransformer();
t.setOutputProperty(OutputKeys.DOCTYPE_SYSTEM, doc.getDoctype().getSystemId());
t.transform(new DOMSource(doc), new StreamResult(out));
} catch(TransformerException e) {
throw new AssertionError(e); // Can’t happen!

Don't Violate the Principle of Least Astonishment

User of API should not be surprised by behavior
_ It's worth extra implementation effort
_ It's even worth reduced performance
public class Thread implements Runnable {
// Tests whether current thread has been interrupted.
// Clears the interrupted status of current thread.
public static boolean interrupted();

Fail Fast–Report Errors as Soon as Possible After They Occur

Compile time is best - static typing, generics
At runtime, first bad method invocation is best
_ Method should be failure-atomic
// A Properties instance maps strings to strings
public class Properties extends Hashtable {
public Object put(Object key, Object value);
// Throws ClassCastException if this properties
// contains any keys or values that are not strings
public void save(OutputStream out, String comments);

Provide Programmatic Access to All Data Available in String Form

Otherwise, clients will parse strings
_ Painful for clients
_ Worse, turns string format into de facto API
public class Throwable {
public void printStackTrace(PrintStream s);
public StackTraceElement[] getStackTrace(); // Since 1.4
public final class StackTraceElement {
public String getFileName();
public int getLineNumber();
public String getClassName();
public String getMethodName();
public boolean isNativeMethod();

Overload With Care

Avoid ambiguous overloadings
_ Multiple overloadings applicable to same actuals
Conservative: no two with same number of args
Just because you can doesn't mean you should
_Often better to use a different name
If you must provide ambiguous overloadings, ensure same behavior for same arguments
public TreeSet(Collection c); // Ignores order
public TreeSet(SortedSet s); // Respects order

Use Appropriate Parameter and Return Types

Favor interface types over classes for input
_Provides flexibility, performance
Use most specific possible input parameter type
_ Moves error from runtime to compile time
Don't use string if a better type exists
_ Strings are cumbersome, error-prone, and slow
Don't use floating point for monetary values
_ Binary floating point causes inexact results!
Use double (64 bits) rather than float (32 bits)
_ Precision loss is real, performance loss negligible

Use Consistent Parameter Ordering Across Methods

Especially important if parameter types identical
#include <string.h>
char *strcpy (char *dest, char *src);
void bcopy (void *src, void *dst, int n);
java.util.Collections – first parameter always
collection to be modified or queried
java.util.concurrent – time always specified as
long delay, TimeUnit unit

Avoid Long Parameter Lists

Three or fewer parameters is ideal
_ More and users will have to refer to docs
• Long lists of identically typed params harmful
_ Programmers transpose parameters by mistake
_ Programs still compile, run, but misbehave!
Two techniques for shortening parameter lists
_ Break up method
_ Create helper class to hold parameters
// Eleven parameters including four consecutive ints
HWND CreateWindow(LPCTSTR lpClassName, LPCTSTR lpWindowName,
DWORD dwStyle, int x, int y, int nWidth, int nHeight,
HWND hWndParent, HMENU hMenu, HINSTANCE hInstance,
LPVOID lpParam);

Avoid Return Values that Demand Exceptional Processing

return zero-length array or empty collection, not null

package java.awt.image;
public interface BufferedImageOp {
// Returns the rendering hints for this operation,
// or null if no hints have been set.
public RenderingHints getRenderingHints();

Exception Design

Throw Exceptions to Indicate Exceptional Conditions

Don’t force client to use exceptions for control flow
private byte[] a = new byte[BUF_SIZE];
void processBuffer (ByteBuffer buf) {
try {
while (true) {
processBytes(tmp, BUF_SIZE);
} catch (BufferUnderflowException e) {
int remaining = buf.remaining();
buf.get(a, 0, remaining);
processBytes(bufArray, remaining);
Conversely, don’t fail silently
ThreadGroup.enumerate(Thread[] list)

Favor Unchecked Exceptions

Checked – client must take recovery action
Unchecked – programming error
Overuse of checked exceptions causes boilerplate

try {
Foo f = (Foo) super.clone();
} catch (CloneNotSupportedException e) {
// This can't happen, since we’re Cloneable
throw new AssertionError();

Include Failure-Capture Information in Exceptions

Allows diagnosis and repair or recovery
For unchecked exceptions, message suffices
For checked exceptions, provide accessors

Refactoring API Designs

Sublist Operations in Vector

public class Vector {
public int indexOf(Object elem, int index);
public int lastIndexOf(Object elem, int index);
Not very powerful - supports only search
Hard too use without documentation

Sublist Operations Refactored

public interface List {
List subList(int fromIndex, int toIndex);
Extremely powerful - supports all operations
Use of interface reduces conceptual weight
_ High power-to-weight ratio
Easy to use without documentation

Thread-Local Variables

// Broken - inappropriate use of String as capability.
// Keys constitute a shared global namespace.
public class ThreadLocal {
private ThreadLocal() { } // Non-instantiable
// Sets current thread’s value for named variable.
public static void set(String key, Object value);
// Returns current thread’s value for named variable.
public static Object get(String key);

Thread-Local Variables Refactored (1)

public class ThreadLocal {
private ThreadLocal() { } // Noninstantiable
public static class Key { Key() { } }
// Generates a unique, unforgeable key
public static Key getKey() { return new Key(); }
public static void set(Key key, Object value);
public static Object get(Key key);
Works, but requires boilerplate code to use
static ThreadLocal.Key serialNumberKey = ThreadLocal.getKey();
ThreadLocal.set(serialNumberKey, nextSerialNumber());

Thread-Local Variables Refactored (2)

public class ThreadLocal {
public ThreadLocal() { }
public void set(Object value);
public Object get();
Removes clutter from API and client code
static ThreadLocal serialNumber = new ThreadLocal();


API design is a noble and rewarding craft
_ Improves the lot of programmers, end-users, companies
This talk covered some heuristics of the craft
_ Don't adhere to them slavishly, but...
_Don't violate them without good reason
API design is tough
_ Not a solitary activity
Perfection is unachievable, but try anyway

个人分类: 写代码的艺术 java
所属专栏: 写代码的艺术
想对作者说点什么? 我来说一句



Google首席工程师Joshua Bloch谈如何设计优秀的API