Open GoogleCodeExporter opened 8 years ago
Kryo2 initially separated object creation from deserialization. There was
Serializer#create to return a new instance and then Serializer#read to populate
that instance. This proved awkward for immutable instances and was confusing
for users. Eventually it was changed to what we have now, which is just
Serializer#read that creates and populates as needed.
While I agree they would be very useful, I don't see a clean way to implement
copyInto and readInto. Separating create and populate proved to be messy, so
going back to that doesn't have much appeal. I hesitate to add new methods to
Serializer, since it is subclassed so many times. Also I try to keep the stack
calls to a minimum, since the stack limits the depth of object graphs that can
be handled, though this last issue is minor.
Possibly using Kryo in Externalizable could be done differently.
FieldSerializer could be subclassed to allow the object instance to be set,
which it would return in FieldSerializer#create, instead of creating a new
instance. In readExternal you'd have to get an instance to Kryo somehow (maybe
through an ObjectInput implementation?), set the FieldSerializer instance to
"this", and then let Kryo do the deserialization. This sounds a bit messy, but
possibly could be streamlined with a nice API.
Original comment by nathan.s...@gmail.com
on 11 Jul 2012 at 3:31
Hi Nate,
I had already implemented a proof of concept in my Library/Project exactly
how you say, in order to use the benefits of Kryo but inside "standard"
Serialization and Clonation. Writing and reading to a persistent JMS queue
through a standard JMS API, now really flyes!!
What I've done:
1.- Created a XSerializer interface:
public interface XSerializer<T extends Object> {
public void readInto(Kryo kryo, Input input, T destination);
}
2.- Create "XFieldSerializer extends FieldSerializer implements XSerializer"
and implement that method copy-and-pasting from FieldSerializer code :-((((
3.- Create another custom-List-serializer of my own (needed for a special
kind of list in my Project), whick also implements XSerializer
4.- When creating Kryo instances, register XFieldSerializer.class as the new
default serializer
5.- Created ObjectOutputWrapperStream and ObjectInputWrapperStream clases
which wrap/delegate tu ObjectOutput / ObjectInput, and set these stream
wrappers as the underlying streams of the Output and Input instances used.
6.- My classes now "implements Externalizable, Cloneable", and uses a
thread-local instance of Kryo in writeExternal, readExternal, and clone.
The defensive trick in readExternal is to ensure that the serializer chosen
by Kryo implements XSerializer, otherwise forcing to use the new "default"
XFieldSerializer.
Surely this could be made more elegant.
This is only optimal as long as the main object in the
serialization/clonation graph is one handled by these new XSerializers
created. For now this is sufficient for my concrete case...
In other case, I'm afraid I will have to make my eXtensions to the rest of
the serializers...
Ciao
-----Mensaje original-----
De: kryo@googlecode.com [mailto:kryo@googlecode.com]
Enviado el: mi�rcoles, 11 de julio de 2012 17:32
Para: serverperformance@gmail.com
Asunto: Re: Issue 79 in kryo: Additions to the API: Add copyInto, readInto
to Serializer
Original comment by serverpe...@gmail.com
on 11 Jul 2012 at 4:08
You shouldn't have to copy paste code from FieldSerializer.
FieldSerializer#read looks like:
public T read (Kryo kryo, Input input, Class<T> type) {
T object = create(kryo, input, type);
...
So you can just write something like:
Object destination;
protected T create (Kryo kryo, Input input, Class<T> type) {
return destination;
}
public void readInto(Kryo kryo, Input input, T destination) {
this.destination = destination;
return read(kryo, input, null);
}
Other "generic" serializers (CollectionSerializer, MapSerializer, etc) should
have a similar create method.
You could avoid the thread local look up in read/writeExternal and clone by
keeping the Kryo reference in your ObjectInput/Output wrappers.
I'm open to more ideas on how we could make using Kryo with Java serialization
cleaner. I have a feeling that there isn't a good way to integrate it nicely,
so hopefully we can make your current approach cleaner.
Original comment by nathan.s...@gmail.com
on 11 Jul 2012 at 4:30
Thank you for the quick response!
This would prevent using the same Kryo instance / serializers for a
standard/direct Kryo usage (in my Project I am using Kryo directly whenever
posible, but integrating with Java serialization (mainly for JMS and RMI
integration).
Or maybe I am losing something...
by keeping the Kryo reference in your ObjectInput/Output wrappers.
Good catch, I'll try it :-)
Original comment by serverpe...@gmail.com
on 11 Jul 2012 at 4:43
Hmm, in that case maybe try:
Object destination;
protected T create (Kryo kryo, Input input, Class<T> type) {
return destination != null ? destination : super.create(kryo, input, type);
}
public void readInto(Kryo kryo, Input input, T destination) {
this.destination = destination;
read(kryo, input, null);
this.destination = null;
}
Original comment by nathan.s...@gmail.com
on 11 Jul 2012 at 4:56
I will try it (but tomorrow... :-)
Original comment by serverpe...@gmail.com
on 11 Jul 2012 at 5:44
It work :-)
A litte bit dirty, and still depends on the actual implementation detalis of
FieldSerializer (read invokes create, copy invokes createCopy) but better than
before.
Please assure that future evolutions doesn't break this compatibility (if
possible, of course)
Thanks!
Original comment by serverpe...@gmail.com
on 12 Jul 2012 at 9:13
Can you pleas add the createCopy method in CollectionSerializer (and use it
inside copy, of course). It exists in ohter generic serializers:
FieldSerializer and MapSerializer (I have had no time to search in others)
Tnx!
Original comment by serverpe...@gmail.com
on 12 Jul 2012 at 1:04
Hi again. One question: how can I replace default serializers? I mean, for
instance, for replacing the default serializer for collection, I'm invoking:
kryoInstance.addDefaultSerializer(Collection.class, XCollectionSerializer.class);
But, looking at the code, it seems that Kryo.getDefaultSerializer searches in
the defaultSerializers list in order starting from 0. So... it won't find my
replacement as there as the default one is found before. True?
How can I do this replacements?
Original comment by serverpe...@gmail.com
on 12 Jul 2012 at 1:50
Added CollectionSerializer#createCopy.
addDefaultSerializer always puts your additions *before* the default default
serializers, so it will find yours first. This may be a little problem, because
order is important. If you add a default serializer for Collection.class, it
goes first, which means it is before Collections.EMPTY_LIST.getClass() and the
other Collections that are there by default. I guess I'll wait until someone
really has the need for full control of the list, I hate to add more APIs when
it isn't needed. Of course you can just copy paste the addDefaultSerializer
calls from the Kryo class and customize them.
Original comment by nathan.s...@gmail.com
on 12 Jul 2012 at 5:38
Original issue reported on code.google.com by
serverpe...@gmail.com
on 8 Jul 2012 at 8:26