Here the list of modifications:
- The architecture based on JMS didn't support thousands of process instances in receive state (tested with 50000 process), it caused high memory consumption (25GB) and several minutes waiting jboss to start to load all jms receivers (about 20 minutes on a server with 12 cores and Oracle 10g). We removed JMS parts and altered to query the database when receiving a message to continue a process . This caused the loss of asynchronous messages which are not necessary for us. Another solution will be required if someone needs to use of asynchronous messages.
- Performance was degradated when executed hundreds of processes at same time causing crashes on JBoss. A ThreadPoolExecutor was added to SoapHandler.java to limit the number of processes executing at same time.
- Added a class to provide a central configuration for JBPM (ConfigurationManager.java), this classes uses Jgroups to replicate configuration modifications on runtime.
- Added static map (PROCESSES_IN_EXECUTION on JbpmContext.java) to provide a way to the cluster doesn't execute the same process at the same time. JGroups was added to support sending of messages between the machines to provide information about which processes are running.
- Added support for 2nd level cache cluster using EhCache and Jgroups, configuration on jbpm.hibernate.cfg.xml and ehcache_jbpm.xml
- Inclusion of table JBPM_NODE_EXECUTED, to track the path a process instance followed on bpel graph, we tried to use JBPM_TOKEN but it's unpredictable and confuse to track a path.
- XMLs messages with quotes ( " and ' ) has several issues when parsing to and from xerces, the quotes are lost due to the fact that XMLUtil.java encodes & to &amp;. This happens with other characters (>, <, space, -, & and #). It was necessary to modify ElementType.java to properly encode/unencode the xml files.
- Added support to XSL transformations using javax.xml.transform.Transformer in org.jbpm.bpel.sublang.xpath.cpqd.XSLTransformation file. The deploy of processes was altered to include xsd files on table JBPM_BYTEBLOCK
- JBPM_JOB has a design problem, it executes only one process by cycle, altered to execute as many as necessary, each one in a transaction with crash recovery. Also added a class to monitor
- Use of JBPM_LOG was altered, studies of database growth showed that this table was responsible for 30% of all database, so a configuration was added to choose to log or not to log, see DbLoggingService.java
- If a process crashes , the transaction is rolled back depending on exception type so all the graph is lost and if the process passed for a node that cannot be executed again the message can't be resended. StartInstance and ResumeInstance was rewritten to provide means to always save data.
- Added Xpath function to send email based on a xml with the data (org.jbpm.bpel.sublang.xpath.cpqd.SendEmail)
- When assigning a node that has several sub nodes with the same name, only the first node is assigned. XPathEvaluator.java (method selectOrCreateNodes) was fixed.
- When copying xmls nodes between different xmls namespace, the original namespaces are copied also. XPathEvaluator.java was fixed.
- LockMonitorThread.java monitors when a job is executing for a long time, but was not full implemented or used, it was modified to properly monitor the jobs.
- ByteBlockChopper.java had performance issues when glueing the chops because it created a new array for each chop, altered to create only one initial array and glue the chops in the right position.
- Added class JbossWebDescriptorTool.java on JBPM-BPEL-TOOLS, this class creates a jboss-web.xml to provide one context (/jbpmservices) for all processes deployed.
- The .war file deployed does not have version information, this causes the old processes instances crashs, several classes of JBPM-BPEL-TOOLS was modified to provide version information on classes generated and file names.
- On deploying a process definition create classes without using namespace for packaging, it causes a crash because of confliting xml nodes with the same name, modified to create packages for each namespace.
- Added column STATE_ to JBPM_PROCESSDEFINITION table, the definition is put in a activing state when deploying, so a process can be deployed with jbpm in use and the users can not access it until .war file is fully deployed.
- CatalogEntry is cached when used for the first time, so an external web service address could not be modified on runtime. The class was modified to support notification of urls for all the process using the addresses on ServiceCatalog.java, even in cluster.
1 of 1 people found this helpful
Thank you for reporting about your usage of jBPM and of the modifications you made. I think this post better suites the development forum. However AFAIK jBPM BPEL (as part of jBPM 3.x) is not developed actively anymore. Support for BPEL has been dropped in jBPM 4.x. If you are in need of a BPEL 2.0 engine than there is JBoss Riftsaw (based upon Apache ODE): [url]http://www.jboss.org/riftsaw[/url]
Thanks a lot Sebastian, i've started a thread there!