Question

I’m working with an application that uses JBPM 3.1 and MySQL. The core problem is that there are processes instances with variables that contain an older version of an external, non-JBPM Serializable class. When the main application is upgraded, these processes instances cause an exception to be thrown by JBPM since the SUID of a specific class instance has changed in the main application.

I believe I have a method for fixing the deserialization process using the technique described in the following:

How to deserialize an object persisted in a db now when the object has different serialVersionUID

However, my problem is figuring out where in MySQL JBPM stores process instance variables, so I can write a program that can interate over all the variables for all instances, an reserialize the variables so the offending class will have the new SUID, so JBPM can operate against the processes.

My initial looking at the JBPM tables, it appears that the JBPM_BYTEARRAY and/or JBPM_BYTEBLOCK may be the tables to operate against. However, I’m unsure how to proceed. I’m guessing each process variable is stored in a wrapping container class. Is that class org.jbpm.context.exe.VariableInstance? Or is it something else?

I figure if I have the proper jar files in the class path, and I know what the main class instance is that JBPM uses to store process variables in MySQL, I can deserialize the class (which will fix the SUID problem with the embedded problem class instance), and reserialize the class back. Since JBPM documentation does mention stuff about converters, I’m unsure if I have to replicate the conversion process JPBM does when deserializing, or if standard java deserialization is enough.

Answer 1

Some analysis of JBPM indicates that binary data may be split across multiple records. This may not be the case for mysql itself, but the JPBM code is written to support multiple RDBMs, and some have limits on the size of binary records.

Since the question earned me a tumbleweed reward, I was not going to get a usable mysql-based answer in within the deadline I had to meet, so I re-considered the core problem and the operating context the problem occurs, and came up with a solution that avoided the needed to perform direct mysql operations.

The main application in question already has some customize modifications to JBPM, so the solution I implemented altered JBPM source which performs the deserialization of process instance variables. This avoids the need to deal with JBPM logic that extracts the deserialized binary data from the RDBMs.

In the class org.jbpm.context.exe.converter.SerializableToByteArrayConverter, I modifed the code to use a custom ObjectInputStream class that returns the latest SUID of a class. The technique of just replacing the descriptor with the latest version of the class as described in the post referenced in the question does not work if the new class includes new fields. Doing so causes an end-of-data exception since the base deserialization code tries to access the “new” fields in the old, deserialized version of the class.

Therefore, I just need to replace the SUID, but keep all other parts of the descriptor the same. Since the JDK does not make ObjectStreamClass extensible, I created a sub-class of ObjectInputStream that returns the new SUID based upon a given calling pattern the java library executes against ObjectInputStream when deserialzing data.

The pattern: When reading the header of a deserialized object, the readUTF() function is called (to obtain the class name) followed by a readLong() call. Therefore, if this calling sequence occurs, and if the readUTF() returned the class name I want to change the SUID of, I return the newer SUID in the readLong() call.

The custom code reads a configuration file that specifies class names and associated SUIDs that should be mapped to the latest SUIDs for the classes listed. This allows mapping of alternate classes in the future w/o modifying the custom code.

Note, this approach is applicable to general deserialization operations, where one needs to map old SUIDs to the latest SUIDs of specified classes, and leaving the other parts of the serialized class descriptor alone to avoid end-of-data problems if the newer class definition includes additional field declarations not present in the older class definition.