[Cayenne-dev] Re: [jira] [Commented] (CAY-1622) Generated classes shouldn't produce serialVersionUID compiler warning

I find that the cgen task does not regenerate the files unless the cayenne model has an updated datetime stamp. However, your point is taken that this will then cause ALL entities to get a new serial number, even if only one changed.

Ari

So yeah, IMO it does have to be like hashcode of sorts based on ObjEntity state...

Andrus

Just committed fix to the trunk (r1178712). UID generation algorithm
uses same approach as java.io.ObjectStreamClass.getSerialVersionUID().
Generated id depends on ObjEntity's superclass, ObjAttributes and
ObjRelationships names and types.

Here is the diff:

Modified: cayenne/main/trunk/framework/cayenne-jdk1.5-unpublished/src/main/java/org/apache/cayenne/map/ObjEntity.java
URL:http://svn.apache.org/viewvc/cayenne/main/trunk/framework/cayenne-jdk1.5-unpublished/src/main/java/org/apache/cayenne/map/ObjEntity.java?rev=1178712&r1=1178711&r2=1178712&view=diff
==============================================================================
--- cayenne/main/trunk/framework/cayenne-jdk1.5-unpublished/src/main/java/org/apache/cayenne/map/ObjEntity.java (original)
+++ cayenne/main/trunk/framework/cayenne-jdk1.5-unpublished/src/main/java/org/apache/cayenne/map/ObjEntity.java Tue Oct 4 07:57:40 2011
@@ -19,6 +19,9 @@

package org.apache.cayenne.map;

+import java.io.ByteArrayOutputStream;
+import java.io.DataOutputStream;
+import java.security.MessageDigest;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collection;
@@ -1250,5 +1253,42 @@ public class ObjEntity extends Entity im
public void setExcludingSuperclassListeners(boolean excludingSuperclassListeners) {
this.excludingSuperclassListeners = excludingSuperclassListeners;
}
+
+ /**
+ * Returns unique id for this ObjEntity.
+ *
+ * @since 3.1
+ */
+ public long getSerialVersionUID() throws Exception {
+ ByteArrayOutputStream bout = new ByteArrayOutputStream();
+ DataOutputStream dout = new DataOutputStream(bout);
+
+ dout.writeUTF(getClassName());
+
+ if (getSuperClassName() != null) {
+ dout.writeUTF(getSuperClassName());
+ }
+
+ for (ObjAttribute attr : getAttributes()) {
+ dout.writeUTF(attr.getName());
+ dout.writeUTF(attr.getType());
+ }
+
+ for (ObjRelationship rel : getRelationships()) {
+ dout.writeUTF(rel.getName());
+ dout.writeUTF(rel.getCollectionType() != null
+ ? rel.getCollectionType() : "to-one");
+ }
+
+ dout.flush();
+
+ MessageDigest md = MessageDigest.getInstance("SHA");
+ byte[] hashBytes = md.digest(bout.toByteArray());
+ long hash = 0;
+ for (int i = Math.min(hashBytes.length, 8) - 1; i>= 0; i--) {
+ hash = (hash<< 8) | (hashBytes[i]& 0xFF);
+ }
+ return hash;
+ }

}


Dzmitry

Very cool.

Can we also create some unit tests for this to ensure it works as advertised. E.g. a possible caveat is consistent alphabetic ordering of relationships and attributes (IIRC they are ordered by default, but it is worth creating some tests to see how entity creation affects the result). Also I think in relationships loop we may add a check for the target entity type.

Andrus

Actually... since CayenneDataObject is a Map, do we even care about attributes and relationships? No matter what they are, serialization format should stay the same..

Now we also have ROP where objects are NOT maps. So we need a different algorithm for those , which is more like the one you implemented (+ my suggestion about relationship type ... for to-one)

That means that we need 2 different algorithms used in server and ROP templates.

Andrus

Now everything is pretty clear about ROP client classes. But what
algorithm should we use for server classes UID generation? Suppose it
probably must be somehow related with CayenneDataObject class version...

If the plan is to get the whole serialization stuff right at this time, maybe it's best to go the extra mile and avoid serializing the Map keys.

Prerequisite 1: You need to know the field names, both on the sending and the receiving end.

Prerequisite 2: You need a way to iterate over all field values, in the same order as the field names; or you need a way to access the field values given a field name in String form.

If these are given, you can transfer just the values and do not need to transfer the keys.

Win: Halve the amount of data that needs to be transferred. Give or take a few, depending on circumstances.

Maybe win: If you're taking the format of serialized data into your own hands, it could be made identical for the ROP and non-ROP case. This would make it easy to record serialized data in one environment and replay it in another, which might help folks who are setting up diagnostics and unit tests.

Nonproblem: Now that SerialVersionUID takes care that the field structure and hence the Map keys are the same on both sides, transferring keys is indeed redundant.

Disclaimer: I know next to nothing about the internals of Cayenne, so this may be much work than it's worth.

So, just my 2 cents worth here.

Regards,
Jo

This is a separate task (not that it is not important), so I'd not mix it with serialVersionUID discussion.

I have similar ideas about changing internal structure of CayenneDataObject (using [] instead of a map). This is a big change (and should give a big performance boost not just during serialization, but in general), but that is more of a long term future...

Andrus

That would definitely require some serious performance measurements.

The other data structure I have been thinking about is the objects themselves, and using Method objects to access the fields when by-name access or iterating over all fields is required. (I shamelessly stole that approach from org.apache.commons.beanutils.PropertyUtilsBean, which has to solve similar problems.)

Just a whacko idea I've been playing with, I have no idea whether it will be truly faster. Method#invoke looks like it could be optimizable, but maybe not well enough to actually beat array access plus run-time type cast.
So: more performance testing :-)

Regards,
Jo