feat(java): add set serializer for concurrent set

[MrChang0]

What does this PR do?

we use Sets.newConcurrentHashSet() to create set, with lastest guava version it use ConcurrentHashMapKeySetView and old version it use Collections.newSetFromMap(map).

remove CopyOnWriteArrayListSerializer from native-image.properties. #1614 forget delete.

Related issues

Does this PR introduce any user-facing change?

• Does this PR introduce any public API change?
• Does this PR introduce any binary protocol compatibility change?

Benchmark

[chaokunyang]

Why not just invoke Collections.newSetFromMap(map)

[MrChang0]

Why not just invoke Collections.newSetFromMap(map)

constructor must be a empty map.

[chaokunyang]

I just found that this method should not be used if collection codegen is false.
We can just implement serialization like this:

    @Override
    public void write(MemoryBuffer buffer, Collection value) {
      Object fieldValue = Platform.getObject(value, offset);
      fury.writeRef(buffer, fieldValue);
    }

    @Override
    public Collection read(MemoryBuffer buffer) {
      final Map map = (Map) fury.readRef(buffer);
      return (Collection) Collections.newSetFromMap(map);
    }

Sorry for our document, we didn't make it clear about how to implement a new collection serializer.

Actually, the org.apache.fury.serializer.collection.AbstractCollectionSerializer#write should check codegen is true

[MrChang0]

I will take this way, but Collections.newSetFromMap(map) must use empty map(the check is in constructor method), so I have to use Platform.putObject to set data.
I need learn more about codegen to make sure code is right.

[chaokunyang]

How about:

final Map map = (Map) fury.readRef(buffer);
return Collections.newSetFromMap(map);

[MrChang0]

How about:

final Map map = (Map) fury.readRef(buffer);
return Collections.newSetFromMap(map);

In JDK17

  SetFromMap(Map<E, Boolean> map) {
      if (!map.isEmpty())
          throw new IllegalArgumentException("Map is non-empty");
      m = map;
      s = map.keySet();
  }

that's why I can't do it.

[chaokunyang]

I see, thanks. If only allow users to create set first, then add values.

If so, the serialization should be simplified.

Here is the serialization code:

public boolean add(E e) { return m.put(e, Boolean.TRUE) == null; }

All map values should not be serialized. And map will be empty, so there is not shared reference for this map.

Your previous implementation seems better, only need to set codegen to true. So fury will generate serialization code for this set. In onCollectonWrite, you write map class, and newCollection, you read class and create map from it. You can get serializer for that map, and it must be a AbstractMapSerializer, then you can use AbstractMapSerializer#newMap to create the empty map, and build the set from it.

Current implementation is OK for me too. The default implementation in Fury will serializer whole map too.

We can merge this PR first, and create another PR to optimize it if you like .

[MrChang0]

I prefer to save map classInfo, It looks likes set serializer and less reflection operate. anyway tomorrow is workday, I will finish it.

[MrChang0]

what fuck discussions

[chaokunyang]

LGTM

[chaokunyang]

UpdateCache is unnecessary, next written item will be different types. Other part looks great

[MrChang0]

fine, I careless.