Open GoogleCodeExporter opened 9 years ago
I would not expect it to. In the scenario presented, it is serializing an I,
based only on what it knows about I. And it knows *nothing* about I (it doesn't
have any serialization members). You have supplied a default concrete
implementation to use when the value is null, but that type is non-creatable.
Even if this worked, it would deserialize with N as 0. I can, however, perhaps
fix the handling a *bit* such that it creates an OS, converts it to null, then
applies the (non-existant) serialization rules of I to this new value.
Original comment by marc.gravell
on 13 Jun 2011 at 10:11
Marc, thanks for the details explanation, although I did not fully grasp it. As
the user of the API, I do not quite understand the semantic difference between
m.Add(typeof(I), false).ConstructType = typeof(O);
and
m.Add(typeof(I), false).AddSubType(1, typeof(O));
Except, that there may be many subtypes, but only a single construct type and
that these statements are mutually exclusive.
If I am wrong (and I guess I am), then what is exactly ConstructType for?
Original comment by mark.kha...@gmail.com
on 13 Jun 2011 at 11:50
Maybe I'll rename it if it is confusing (with suitable [Obsolete(...)] etc).
The use-case there is when you want to specify a default implementation of an
interface, most likely because it is interface based but you want to specify
the actual model to use. The serialization is still based entirely off the
interface in this case, it has simply managed to create an object.
The difference with AddSubType is that here the serializer is considering the
subtypes *as active in the serialization*; for example, considering any
serialization members specific to the concrete type (not necessarily specified
on the interface).
Original comment by marc.gravell
on 13 Jun 2011 at 12:24
I can see that AddSubType and ConstructType are not mutually exclusive, meaning
API allows to write something like this:
m.Add(typeof(I), false).AddSubType(1, typeof(O)).ConstructType = typeof(O2);
Is it meaningful?
Original comment by mark.kha...@gmail.com
on 13 Jun 2011 at 12:34
Yes; the `ConstructType` has a use in that when specified it will be used (in
place of `I`) when no sub-type is specified (which is possible considering the
two ends of data need not have exactly the same version, etc). If `I` is an
abstract base-class or interface, or you simply have a preferred default
implementation, then this should allow it to function as preferred.
Original comment by marc.gravell
on 15 Jun 2011 at 10:24
But you are going to serialize 'I' according to the signature of 'I', rather
than its ConstructType, which is different from the Subtypes, where you
serialize according to the particular subtype declaration, am I right?
Original comment by mark.kha...@gmail.com
on 15 Jun 2011 at 10:56
This only applies during deserialization, but yes; it will deserialize
according to `I` rules.
Original comment by marc.gravell
on 15 Jun 2011 at 11:33
May I ask why? After all, you could deserialize according to the
ConstructType, respecting the surrogate logic, of course. Both approaches
are valid, aren't they? May be both should be supported?
Original comment by mark.kha...@gmail.com
on 15 Jun 2011 at 12:15
What data would it be deserializing? If the sender had knowingly serialized
values from the concrete type, it would have to be mapped as a sub-type
(=sub-message), otherwise there is a vast problem of field conflicts. To be
encoded as a sub-message, I need a field-marker (the number in the AddSubType
call). ConstructType is *only* intended for a default implementation of an
otherwise abstract type.
If you need serialization *of that type* (which isn't the expected type, I),
then it will need to be properly defined as a sub-type.
Original comment by marc.gravell
on 15 Jun 2011 at 12:26
I think I understand finally. protobuf-net must maintain the cross platform
nature of the protocol buffers specification. Meaning, the receiving end
must rely exclusively on the proto spec, which mandates that interface
serialized according to a concrete type must have this concrete type
described in the same proto spec and include a field number for it. Hence
ConstructType cannot be used. Is my understanding correct?
Original comment by mark.kha...@gmail.com
on 15 Jun 2011 at 12:34
The google spec *has no notion at all* of interfaces, concrete types, abstract
types, inheritance, etc. I use a few tricks here to provide additional features
*within* the spec, but my bigger point is simply around the intent of
ConstructType. To provide safety from field collisions, I need some sub-message
separation between data of different types (even if for the same object).
ConstructType does not provide a sub-message, as it has no "field". If it *had*
a field, it would be completely indistinguishable from AddSubType - so if that
is the intent, just use AddSubType.
Original comment by marc.gravell
on 15 Jun 2011 at 1:14
Original issue reported on code.google.com by
mark.kha...@gmail.com
on 13 Jun 2011 at 8:57