I'm having some issues trying to serialize/deserialize a complex object graph using protobuf-net.
I'm working on a legacy application and we're using .Net Remoting to connect a GUI client to a C# service. We are seeing poor performance with overseas users due to the serialized size of our object graphs using the default BinaryFormatter
, which is exacerbated by the limited bandwidth in-between the client and server (1Mbit/s).
As a quick win, I thought I'd put together a proof of concept to see if there were any performance gains to be had by using protobuf-net instead, by implementing ISerializable
. As I was testing I ran into an issue whereby object references weren't being maintained.
I've put together an example which repros the issue. I'm expecting that the object in the Dictionary
(Items[1]) and the object B.A will be the same as I've specified AsReference=true
in the ProtoMember
attribute.
Using protobuf-net 2.0.0.619
, I'm seeing an exception thrown when deserializing (A reference-tracked object changed reference during deserialization).
If this isn't a supported scenario the please let me know.
Test
[Test]
public void AreObjectReferencesSameAfterDeserialization()
{
A a = new A();
B b = new B();
b.A = a;
b.Items.Add(1, a);
Assert.AreSame(a, b.A);
Assert.AreSame(b.A, b.Items[1]);
B deserializedB;
using (var stream = new MemoryStream())
{
Serializer.Serialize(stream, b);
stream.Seek(0, SeekOrigin.Begin);
deserializedB = Serializer.Deserialize<B>(stream);
}
Assert.AreSame(deserializedB.A, deserializedB.Items[1]);
}
Class definitions
[Serializable]
[ProtoContract]
public class A
{
}
[Serializable]
[ProtoContract]
public class B
{
[ProtoMember(1, AsReference = true)]
public A A { get; set; }
[ProtoMember(2, AsReference = true)]
public Dictionary<int, A> Items { get; set; }
public B()
{
Items = new Dictionary<int, A>();
}
}
Edit: this should work from the next build onwards simply by marking the type's AsReferenceDefault
:
[ProtoContract(AsReferenceDefault=true)]
public class A
{
// ...
}
At the current time this is sort of an unsupported scenario - at least, via the attributes it is unsupported; basically, the AsReference=true
currently is referring to the KeyValuePair<int,A>
, which doesn't really make sense since KeyValuePair<int,A>
is a value-type (so this can never be treated as a reference; I've added a better message for that in my local copy).
Because KeyValuePair<int,A>
acts (by default) as a tuple, there is currently nowhere to support the AsReference
information, but that is a scenario I would like to support better, and I will be investigating this.
There was also a bug that meant that AsReference
on tuples (even reference-type tuples) was getting out-of-order, but I've fixed that locally; this was where the "changed" message came from.
In theory, the work for me to do this isn't huge; the fundamentals already work, and oddly enough it came up separately on twitter last night too - I guess "dictionary pointing to an object" is a very common scenario. At a guess, I imagince I'll add some atribute to help describe this situation, but you can actually hack around it at the moment using a couple of different routes:
1: configure KeyValuePair<int,A>
manually:
[Test]
public void ExecuteHackedViaFields()
{
// I'm using separate models **only** to keep them clean between tests;
// normally you would use RuntimeTypeModel.Default
var model = TypeModel.Create();
// configure using the fields of KeyValuePair<int,A>
var type = model.Add(typeof(KeyValuePair<int, A>), false);
type.Add(1, "key");
type.AddField(2, "value").AsReference = true;
// or just remove AsReference on Items
model[typeof(B)][2].AsReference = false;
Execute(model);
}
I don't like this much, because it exploits implementation details of KeyValuePair<,>
(the private fields), and may not work between .NET versions. I would prefer to replace KeyValuePair<,>
on the fly via a surrogate:
[Test]
public void ExecuteHackedViaSurrogate()
{
// I'm using separate models **only** to keep them clean between tests;
// normally you would use RuntimeTypeModel.Default
var model = TypeModel.Create();
// or just remove AsReference on Items
model[typeof(B)][2].AsReference = false;
// this is the evil bit: configure a surrogate for KeyValuePair<int,A>
model[typeof(KeyValuePair<int, A>)].SetSurrogate(typeof(RefPair<int, A>));
Execute(model);
}
[ProtoContract]
public struct RefPair<TKey,TValue> {
[ProtoMember(1)]
public TKey Key {get; private set;}
[ProtoMember(2, AsReference = true)]
public TValue Value {get; private set;}
public RefPair(TKey key, TValue value) : this() {
Key = key;
Value = value;
}
public static implicit operator KeyValuePair<TKey,TValue>
(RefPair<TKey,TValue> val)
{
return new KeyValuePair<TKey,TValue>(val.Key, val.Value);
}
public static implicit operator RefPair<TKey,TValue>
(KeyValuePair<TKey,TValue> val)
{
return new RefPair<TKey,TValue>(val.Key, val.Value);
}
}
This configures something to use instead of KeyValuePair<int,A>
(converted via the operators).
In both of these, Execute
is just:
private void Execute(TypeModel model)
{
A a = new A();
B b = new B();
b.A = a;
b.Items.Add(1, a);
Assert.AreSame(a, b.A);
Assert.AreSame(b.A, b.Items[1]);
B deserializedB = (B)model.DeepClone(b);
Assert.AreSame(deserializedB.A, deserializedB.Items[1]);
}
I do, however, want to add direct support. The good thing about both of the above is that when I get time to do that, you just have to remove the custom configuration code.
For completeness, if your code is using Serializer.*
methods, then rather than create / configure a new model, you should configure the default model:
RuntimeTypeModel.Default.Add(...); // etc
Serializer.*
is basically a short-cut to RuntimeTypeModel.Default.*
.
Finally: you should not create a new TypeModel
per call; that would hurt prerformance. You should create and configure one model instance, and re-use it lots. Or just use the default model.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With