问题
In one of my projects, I have two \"data transfer objects\" RecordType1 and RecordType2 that inherit from an abstract class of RecordType.
I want both RecordType objects to be processed by the same RecordProcessor class within a \"process\" method. My first thought was to create a generic process method that delegates to two specific process methods as follows:
public RecordType process(RecordType record){
if (record instanceof RecordType1)
return process((RecordType1) record);
else if (record instanceof RecordType2)
return process((RecordType2) record);
throw new IllegalArgumentException(record);
}
public RecordType1 process(RecordType1 record){
// Specific processing for Record Type 1
}
public RecordType2 process(RecordType2 record){
// Specific processing for Record Type 2
}
I\'ve read that Scott Meyers writes the following in Effective C++ :
\"Anytime you find yourself writing code of the form \'if the object is of type T1, then do something, but if it\'s of type T2, then do something else,\' slap yourself.\"
If he\'s correct, clearly I should be slapping myself. I don\'t really see how this is bad design (unless of course somebody subclasses RecordType and adds in a RecordType3 without adding another line to the generic \"Process\" method that handles it, thus creating a NPE), and the alternatives I can think of involve putting the brunt of the specific processing logic within the RecordType classes themselves, which really doesn\'t make much sense to me since there can in theory be many different types of processing I\'d like to perform on these records.
Can someone explain why this might be considered bad design and provide some sort of alternative that still gives the responsibility for processing these records to a \"Processing\" class?
UPDATE:
- Changed
return nulltothrow new IllegalArgumentException(record); - Just to clarify, there are three reasons a simple RecordType.process() method would not suffice: First, the processing is really too far removed from RecordType to deserve its own method in the RecordType subclasses. Also, there are a whole slew of different types of processing that could theoretically be performed by different processors. Finally, RecordType is designed to be a simple DTO class with minimal state-changing methods defined within.
回答1:
The Visitor pattern is typically used in such cases. Although the code is a bit more complicated, but after adding a new RecordType subclass you have to implement the logic everywhere, as it won't compile otherwise. With instanceof all over the place it is very easy to miss one or two places.
Example:
public abstract class RecordType {
public abstract <T> T accept(RecordTypeVisitor<T> visitor);
}
public interface RecordTypeVisitor<T> {
T visitOne(RecordType1 recordType);
T visitTwo(RecordType2 recordType);
}
public class RecordType1 extends RecordType {
public <T> T accept(RecordTypeVisitor<T> visitor) {
return visitor.visitOne(this);
}
}
public class RecordType2 extends RecordType {
public <T> T accept(RecordTypeVisitor<T> visitor) {
return visitor.visitTwo(this);
}
}
Usage (note the generic return type):
String result = record.accept(new RecordTypeVisitor<String>() {
String visitOne(RecordType1 recordType) {
//processing of RecordType1
return "Jeden";
}
String visitTwo(RecordType2 recordType) {
//processing of RecordType2
return "Dwa";
}
});
Also I would recommend throwing an exception:
throw new IllegalArgumentException(record);
instead of returning null when neither type is found.
回答2:
My suggestion:
public RecordType process(RecordType record){
return record.process();
}
public class RecordType
{
public RecordType process()
{
return null;
}
}
public class RecordType1 extends RecordType
{
@Override
public RecordType process()
{
...
}
}
public class RecordType2 extends RecordType
{
@Override
public RecordType process()
{
...
}
}
If the code you need to execute is coupled to something the model shouldn't know (like UI) then you will need to use a kind of double dispatch or visitor pattern.
http://en.wikipedia.org/wiki/Double_dispatch
回答3:
Another possible approach would be to make process() (or perhaps call it "doSubclassProcess()" if that clarifies things) abstract (in RecordType), and have the actual implementations in the subclasses. e.g.
class RecordType {
protected abstract RecordType doSubclassProcess(RecordType rt);
public process(RecordType rt) {
// you can do any prelim or common processing here
// ...
// now do subclass specific stuff...
return doSubclassProcess(rt);
}
}
class RecordType1 extends RecordType {
protected RecordType1 doSubclassProcess(RecordType RT) {
// need a cast, but you are pretty sure it is safe here
RecordType1 rt1 = (RecordType1) rt;
// now do what you want to rt
return rt1;
}
}
Watch out for a couple of typos - think I fixed them all.
回答4:
Design is a means to an end, and not knowing your goal or constraints, nobody can tell whether your design is good in that particular situation, or how it might be improved.
However, in object oriented design, the standard approach to keep the method implemention in a separate class while still having a separate implementation for each type is the visitor pattern.
PS: In a code review, I'd flag return null, because it might propagate bugs rather than reporting them. Consider:
RecordType processed = process(new RecordType3());
// many hours later, in a different part of the program
processed.getX(); // "Why is this null sometimes??"
Put differently, supposedly unreachable code paths should throw an exception rather than result in undefined behaviour.
回答5:
Bad design in one think, as in your example, not using visitor pattern, when applicable.
Another is efficiency. instanceof is quite slow, compared to other techniques, such as comparing against class object using equality.
When using visitor pattern, usually an effective and elegant solution is using Map for mapping between support class and Visitor instance. Large if ... else block with instanceof checks would be very ineffective.
来源：https://stackoverflow.com/questions/8841577/is-this-use-of-the-instanceof-operator-considered-bad-design