I think you've almost answered your own question there.
The second piece of code is preferred over the first because it offers type-safety. With the first piece, if you have a similar enumeration of Fish then you can say something like
MostNotableGrievance grievance = Fish.Haddock;
and the compiler won't care. If Fish.Haddock = 2 then the above will be exactly equivalent to
MostNotableGrievance grievance = MostNotableGrievance.FlipFlopping;
but obviously not instantly readable as such.
The reason enumerations are often no better is because the implicit conversion to and from int leave you with the exact same problem.
But, in C#, there is no such implicit conversion, so an enum is a better structure to use. You'll rarely see either of the first two approaches used.
I fully realize that I am 2 years late to the party :-) but probably few clarifications will help other devs.
What the other answers missed to address was the mentioning of the "Kryonet" which adds additional twist to the problem: usually when kryonet is used it means that
a) the client is an android app and
b) the client is some game
If both conditions are met IMHO using the instanceof if-else sequence is the way to go.
The usual way of structuring client/server app in our case android game + "pure" java server is to use (at least) 3 projects:
- "common" - which contains code common for the server and the client, most importantly all the message classes (which usually implement some tag/marker interface as
GameMessage
).
- server - contains the server and includes the
common
project. Usually all messages are simply registered in Kryonet, i.e. no pooling is used (because on server GC is not considered a problem)
- android client - contains the android app (game) and includes the
common
project. Because in java games creating objects during the game may trigger the GC it is considered the biggest mortal sin (which leads straight to the hell of missed frames) it is considered good practice1 to use object pooling for the game messages (this requires little customization on the Kryonet/kryo in order to use object pools). All messages are (must be) registered in kryonet in the same order as on the server in order things to work.
First, lets look at the alternatives and see why they are not fit for the structure/implementation described above:
1. They say 'Use polymorphism and method overriding'
Take a look here for an example of the proposed solution (accepted answer by jmg). It looks very elegant but there there are at least 2 big problems:
You can't just use parameterless overrided method like do()
because in real implementation you will some need external data to act upon. For example when receiving PlayerAbandonedGame
message on the server you will need at least the particular game object as parameter in order to call removePlayer(Player pl)
method. NewGameState
message will require reference to the particular GameState
variable that will be updated and so on. One very ugly solution is overrided method to have single parameter like TheWholeThing
which contains all the objects that may be needed. It is needless to say that this not only violates the principle of encapsulation but also "spreads" the game logic across multiple message classes.
Remember that we have the common
project and all the messages are defined there. Even if server-to-client messages are completely distinct set to client-to-server messages it will be simply stupid to put server or client logic in the common
project.
Game message objects are just data containers. It is up to the server or the client to handle that data in a particular way (or even ignore it as often happens with out of order UDP messages).
2. They say 'Use reflection'
Our context is 'a) the client is an android app'. The reaction of every android developer hearing Use reflection
should be to stand up, clear his/hers throat and with his/hers best british accent to say: 'Are you completely mAd?!' :-)
Reflection is very slow on older Android versions, it is slow even now (5.0). Also IMO using reflection is the ugliest possible solution not just for this particular problem. It should be avoided up in the stack, i.e. in app dev.
Way to go
Back to the original question: if one really insist on using switch
the solution is to have base class like:
public abstract class GameMessage {
byte mCode; // not private because of kryo
public GameMessage(byte code) {
super();
mCode = code;
}
public byte getCode() {
return mCode;
}
}
concrete message will look like:
public class PlayerAbandonedGame extends GameMessage {
// some fields
public PlayerAbandonedGame() {
super((byte) 42);
}
}
and then message handler will look like:
public synchronized void onMessageReceived(GameMessage msg) {
switch(msg.getCode()) {
case 42:
// do something
break;
default:
mLogger.error("Unknown message code {}", msg.getCode());
// may be even throw an exception?
break;
}
// ...
}
Of course it is advisable to use enum values instead of hardcoded numbers...
Potential drawback is (as mentioned in other answers) if you create new message class and forget to add it to some of the switch statements. The compiler cannot catch this problem and you will hit it only runtime (in the default
case).
Real (small) drawback is that code
field is send each time and adds to the traffic. In reality this adds about 1-10% to the size of the messages (if you use byte
for the code field).
How I do it
I personally use if-else sequence with instanceof test like:
public synchronized void onMessageReceived(GameMessage msg) {
if (msg instanceof MyLetterAttemptResult) {
handleMyLetterAttemptResult((MyLetterAttemptResult) msg);
mPools.release(msg);
} else if (msg instanceof InAutoMatchQueue) {
handleInAutoMatchQueue((InAutoMatchQueue) msg);
} else if (msg instanceof AutoMatchRetry) {
handleAutomatchRetry();
// more else-ifs
} else {
mLogger.error("Unknown message {}", msg.getClass().getSimpleName());
// throw exception may be?
}
}
Isn't that "code smell"? You bet it is, but after trying all other alternatives it seems to me that it is best. To reduce the smelliness it is better to try to keep code in the ifs short like above. Just call handle* method and release message back to the pool (if it is pooled).
Isn't that slow? It is a bit slower than the switch(msg.getCode())
but not that much. Here is some benchmarking (answer by Dan Mihai Ile). Advantage is that you don't have the code
field send each time over the network. Keep the most frequently used messages at the top so they are found first.
Isn't that error prone? There is a risk if you create new message class and forget to add corresponding else if
. The thing is that you have just two separate message handling methods - one on the server and one on the client. Both are the single most important code pieces repectively on the server and the client. You will simply have to not mess it up. Just add the new message class immediately to the unit test that tests onMessageReceived()
.
This is not OOP! I have a better solution! I will be very happy to discuss it and use it :-). btw, I wonder if @Cameron (the author of this question) found better solution. After all this question is 2 years old...
1 For me avoiding the GC is more like "utter crap" than "best practice". As Eonil commented in 1: I regard avoiding GC on GC based language is a kind of insane trial.
Best Answer
Both
switch
statements and polymorphism have their use. Note though that a third option exists too (in languages which support function pointers / lambdas, and higher-order functions): mapping the identifiers in question to handler functions. This is available in e.g. C which is not an OO language, and C# which is*, but not (yet) in Java which is OO too*.In some procedural languages (having no polymorphism nor higher-order functions)
switch
/if-else
statements were the only way to solve a class of problems. So many developers, having accustomed to this way of thinking, continued to useswitch
even in OO languages, where polymorphism is often a better solution. This is why it is often recommended to avoid / refactorswitch
statements in favour of polymorphism.At any rate, the best solution is always case dependent. The question is: which option gives you cleaner, more concise, more maintainable code in the long run?
Switch statements can often grow unwieldy, having dozens of cases, making their maintenance hard. Since you have to keep them in a single function, that function can grow huge. If this is the case, you should consider refactoring towards a map based and/or polymorphic solution.
If the same
switch
starts to pop up in multiple places, polymorphism is probably the best option to unify all these cases and simplify code. Especially if more cases are expected to be added in the future; the more places you need to update each time, the more possibilities for errors. However, often the individual case handlers are so simple, or there are so many of them, or they are so interrelated, that refactoring them into a full polymorphic class hierarchy is overkill, or results in a lot of duplicated code and/or tangled, hard to maintain class hierarchy. In this case, it may be simpler to use functions / lambdas instead (if your language allows you).However, if you have a
switch
in a single place, with only a few cases doing something simple, it may well be the best solution to leave it like it is.*I use the term "OO" loosely here; I am not interested in conceptual debates over what is "real" or "pure" OO.