Here’s an issue I’ve been wrestling with and cannot come up with a solid solution:
Should IDs be included inside structures, or is that metadata that should be encapsulated at a higher level?

Here is a more concrete example. We are tracking the state of a given Fighter in the game. This is a stateless server environment, so we need to transfer as little information as possible, but we must still describe everything. We assume there is a larger datastore on the server that has additional constant “definition” information for each Fighter.

Here is what the definition looks like for the external metadata version. Incidentally we are using Thrift, but the issue is the same in any environment.

// current state for fighter.
struct FighterState {
1: required i32 CurHealth;
2: required i32 MaxHealth;
}

struct FighterList {
// indexing
1: required map<FighterID,FighterState> fighters;

// connects a fighter to the definition datastore.
2: required map<FighterID, string> definitionIndex;
}

Immutable information is clearly separated from mutable. However, every method that needs to operate on a Fighter requires two parameters: The FighterList and the FighterID to identify which one is involved.

So this is what the information looks like if we encapsulate it all in one location.

// current state for fighter.
struct FighterState {
// indexing. this is immutable.
1: required FighterID FighterID;
2: required string DefinitionID;

// actual state
3: required i32 CurHealth;
4: required i32 MaxHealth;
}

This produces a mixed-mutability object. This feels less than ideal. Every time I transfer the state between client and server, I repeat the DefinitionID. However, if I want to pass a Fighter to a method, I can simply pass the FighterState and have enough information.

So let’s approach this in another way. What if we use compositing?

// current state for fighter. This is mutable.
struct FighterState {
1: required i32 CurHealth;
2: required i32 MaxHealth;
}

// immutable
struct FighterInfo {
1: required FighterID FighterID;
2: required string DefinitionID;
}

// combine them for runtime use
struct Fighter {
1: required FighterState State;
2: required FighterInfo Info;
}

We have now clearly separated the mutable from the immutable. We also can pass just the one Fighter to methods. But this feels very wordy. Also, the needs of the client and the server are slightly different. The server needs the complete “view” of the fighter definition, while the client can use a trimmed down “view”.

This can be simplified by merging the FighterInfo into the Fighter. Not quite as clean of a separation, but it maintains the mutability model.


// current state for fighter. This is mutable.
struct FighterState {
1: required i32 CurHealth;
2: required i32 MaxHealth;
}

// Network view of the fighter. It includes a mutable base, but the additional fields are immutable.
// This is minimal information for sending across the network.
struct FighterNetworkView {
1: required FighterID FighterID;
2: required string DefinitionID;
3: required FighterState State;
}

// the server is now able to add additional information – the complete definition for the fighter.
// This no longer must be handled in Thrift, it can be a subclass of FighterNetworkView, which simplifies accessing.
struct FighterServerView {
1: required FighterNetworkView Network;
2: required FighterDefinition Definition; // we no longer need to look this up externally
}

Now it is implied that only the base FighterState is mutable, while preserving just one parameter to methods. Additionally, this assists unit testing because there is no longer an external dependency to grab the Definition information at each use.

This still feels slightly wrong however because we are making a FighterServerView technically “mutable” because the base class is.

On the other hand, I’m used to SQL tables where IDs are regularly included with the data rather than external, necessary for lookups.

Opinions anyone?