Hi, A brief update on this topic. After doing a simple test with example https://github.com/apache/incubator-kie-kogito-examples/tree/stable/serverless-workflow-examples/serverless-workflow-data-index-quarkus, the number of updates over Nodes table is n*n, so we manage to obtain a perfect quadratic performance degradation. The problem is worse in the case of Serverless Workflow than in BPMN because we the number of nodes is greater than the number of states. In that example N is 16, but for a more complex workflow it would be certainly large. I think that this is more related to how we are handling JPA in the code, in particular the mapping from model to entity (basically JPA is blind and has to update all nodes for every write because it believes the node has been updated, although it is not) than an issue in the table definition. In fact, when using JPA, separating the server model from the JPA entity is not a good idea, especially if the entity contains collections. I will try to change that without breaking anything.
On Wed, Nov 22, 2023 at 12:10 PM Enrique Gonzalez Martinez < [email protected]> wrote: > After the events split you now will need to create a node instance > model instance of making independent from the process instance. > That should do the trick. > > Regarding deleting/inserting it was fixed at some point. > > El mar, 21 nov 2023 a las 20:22, Francisco Javier Tirado Sarti > (<[email protected]>) escribió: > > > > Hi Martin, > > I have a task to review performance of > > > > ProcessInstanceNodeDataEventMerger > > My idea is to reduce the number of delete inserts when processing events > > and try to do it incremental. > > That should improve performance. > > PS: > > I was planning to send an e-mail tomorrow announcing that in case you > were > > already working on a fix for that. I assume you are not and I would be > > sending a PR soon. > > > > On Tue, Nov 21, 2023 at 6:09 PM Martin Weiler <[email protected]> > > wrote: > > > > > I looked into the new examples using data-index persistence addon - > Neus' > > > PR#1813 [1] for serverless and Pere's branch [2] for workflow (great > job > > > both!) - and they work without issues using single requests. However, > under > > > some load (I used 'ab' for testing with a light concurrency of 10 > parallel > > > requests) I ran into the following problems: > > > > > > (1) Large number of insert/delete calls (eg. for tables such as nodes, > > > definitions, etc) > > > > > > (2) Hibernate OptimisticLockExceptions / StaleStateExceptions > > > > > > (3) DB deadlocks > > > > > > (4) Error responses, slow response times > > > > > > The reason I am reaching out with this topic here is to find out if we > are > > > aware of this issue, and if someone is already looking into or being > > > assigned to it? > > > > > > Thanks, > > > Martin > > > > > > [1] https://github.com/apache/incubator-kie-kogito-examples/pull/1813 > > > [2] > > > > https://github.com/pefernan/kogito-examples/tree/example_data-index_persistence > > > > > > --------------------------------------------------------------------- > > > To unsubscribe, e-mail: [email protected] > > > For additional commands, e-mail: [email protected] > > > > > > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
