Flatbuffers: Why is "inheritance" not supported by FlatBuffers?

[exhau]

"Inheritance" I mean here is simply (no multiple inheritance)

table Monster { .. }
table XxxMonster : Monster {
NewMember1 : type
NewMember2 : type
}

The implementation could be

struct Monster : private flatbuffers::Table {
    VT_POS = 4,
    VT_MANA = 6,
    VT_HP = 8,
    VT_NAME = 10,
    VT_INVENTORY = 14,
    VT_COLOR = 16,
    VT_WEAPONS = 18,
    VT_EQUIPPED_TYPE = 20,
    VT_EQUIPPED = 22
...
}

struct XxxMonster : public Monster {
    VT_NewMember1 = 26
    VT_NewMember2 = 28
...
}

I don't see any reason why there is no such a feature. Is there any concern?

Thanks

[ghost]

The big reason is that the fragile base class problem would be even more problematic here than in other systems. In FlatBuffers you can only add new fields at the end (or as the next higher id), which means that once you inherit, you can never add anything safely to the base class anymore.

[exhau]

I think it can be a feature. One can choose to use or not to use. In my case, I plan to use flatbuffers internally, like in-memory data storage (can be temporarily stored on hard drive).

In current design without inheritance, what is the best practice to represent this relationship?

[promethe42]

Composition ?

table Monster { .. }
table XxxMonster {
  base:Monster;
  ...
}

[exhau]

Thank you for your reply. This is an option, but disadvantage is polymorphism.

Let say there is a buffer from somewhere. I know it is a kind of Monster but don't know the exact type(XxxMonster).
Using the inheritance way, the buffer can be read as either a Monster or a XxxMonster with no problem.
If we don't know the number of inheritance layers, there is no way to read a buffer of derived type as a base type unless inserting type information into buffer.

[promethe42]

table Monster { .. }
table XxxMonster {
  base:Monster;
  ...
}
table YyyMonster {
  base:Monster;
  ...
}
union AnyMonster {
  XxxMonster,
  YyyMonster
}

You can use AnyMonster to store an XxxMonster or an YyyMonster.
FlatBuffers will also store the actual type of the union field so you can cast it to the actual type (cf "Union" in https://google.github.io/flatbuffers/flatbuffers_guide_writing_schema.html).

[xyczero]

@promethe42 This solution may be the best. Otherwise , we have to generate code for your inheritance by manually when deserializing from the hard drive.

[exhau]

Thank you guys.

What I am planning to do is to wrap a buffer as a normal class. When a buffer is read from disk or in-memory storage and binded to a certain type, getters can directly access the buf without copying anything. If we wish to edit it, we can unpack it to an object type(copy/deserialize), do changes and then save the changes. Saving is actually packing the object to byte array again.

However, on the other hand, I also wish the classes can be used as normal ones, with runtime polymorphism.

Inheritance can be complex and unpredictable, for example, B:A, C:B, D:B,E:D... (fortunately no multiple inheritance like C:B,A). In the example, class A and E may belong to different modules developed by two teams. E's module is a standalone addon. The solution by "Union" is not very reasonable in my case. The module containing A can't know E in advance.

Our application has a lot of data managed by Key-Value database, and there are frequent derializations (which are expensive) right now. In most time, they are immutable.
And we don't plan to use flatbuffer for network communication or persisted storage. All buffers are temporary running data. Compared to the concern of fragile base class, it is more important for us that inheritances from different team do not touch framework. In other words, we stick to OO but wish to change the memory layout of an (readonly)object. In our case, if framework changes, everything re-generates the code.

Sorry if I didn't make it clear. I just want to utilize flatbuffer to avoid unnecessary copy and deserialization within the library. Maybe it's a misuse or overkill. Forgive me for my amateurish.

[ghost]

I understand your use case, but it is a non-trivial feature (assuming we support it in all languages), and it is something that most users should not be using, since they'll get themselves into trouble. I think @promethe42 's solution is the best, assuming you need to pass part of a "derived class" to a function that expects a "base class".

[exhau]

@gwvo I understand the situation. Do you think it is doable if I fork the repo and do the change myself?

assuming you need to pass part of a "derived class" to a function that expects a "base class".

I need to pass the _entire_ derived class instead of _a part of_ to a function in base module. And then cast it to a base type. Virtual functions should work.

[ghost]

You can't have virtual methods in the table accessor objects, since they point to inside the FlatBuffer.

You can still have polymorphism through templates (in C++), i.e. any two object that happen to have the same fields can be accessed this way. You can achieve this relatively simply by just having a schema feature that copies all the fields of a base table.

[exhau]

Virtual methods are implemented in the wrapper class. As long as the buffer has the type information to construct correct wrapper class.

Assuming there exists "Inheritance" by appending, one can cast a buffer of derived buffer to any base type buffer without any cost.

To make it clear, the design could be like the following code.

BaseBuffer { int typehash; }   // general base with only type information
MonsterBuffer : BaseBuffer{...}
XxxMonsterBuffer : MonsterBuffer{...}

WrapperBase
{
protected:
BaseBuffer* buf
}

WrapperMonster : WrapperBase
{
public:
   getters;
private:
    MonsterBuffer* TypeCast() { return static_cast<MonsterBuffer*>(buf);}
}

WrapperXxxMonster : WrapperMonster
{
public:
   getters;
private:
    XxxMonsterBuffer* TypeCast() { return static_cast<XxxMonsterBuffer*>(buf);}
}

Given a raw buffer, we can retrieve type information from the very beginning, then use correct constructor (here it is WrapperXxxMonster ). Wrapper classes now can have virtual methods

[grandprixgp]

Are there any full CPP examples of this type of pseudo-inheritance shown by @promethe42 here?

[stale[bot]]

This issue has been automatically marked as stale because it has not had activity for 1 year. It will be automatically closed if no further activity occurs. To keep it open, simply post a new comment. Maintainers will re-open on new activity. Thank you for your contributions.