Problem updating Dynamic Process Definition with Zero Downtime Deployment

[parovi]

Hello there,

We were shortly doing an update of one of our Dynamic Process Definitions (DPD) for order management during a Zero Downtime Deployment (ZDD)
During that time, and for a short period of time (around 10-15 minutes), some of the created Business Processes (BPs) got broken on some random actions. In the ProcessTaskLog was nothing to be seen.

Checking the Logs I could find this message:

java.lang.IllegalStateException: orderManagementProcess(VERSION:2) is not equal to orderManagementProcess(VERSION:1

Taking a deeper look I could confirm that all the BPs that were broken had version 2 assigned, but not all of them were broken.

I also see that in that period BPs with version 1 or 2 were created indistinctly.

These findings brought me to this conclusion:

1. The DPD is being imported to the Database via Impex
2. We are using a Percona MySQL Cluster and the Impex is played first in the master node, and then be replicated to the slave node
3. During this happens, orders are being processed. I'm not sure how the version is assigned for the BP, but it looks like it’s based in the DPD sited in one of the Percona nodes.
4. When the BP then starts, the actions are distributed between different server nodes. And here we have probably the problem:
   a. BP get the new version assigned (in this case, version 2)
   b. The node where the action is executed communicates with a DB node which was still not updated (that means, no version 2 exists there)
   c. Action cannot recognise the version and stop working (without logs)
5. No problems happen when:
   a. BP get the old version
   b. Actions are executed in the node where the new version is present
6. When the DB nodes are synced, the problem is gone

And now my question:

Is there a way to configure the update of the DPD, so that it's working in a ZDD for a clustered environment using Percona?

Note: No ZDD is not a go for our client.

And forgot to mention: we use SAP Commerce version 1905

[parovi]

After diving deeper into the code and reading some documentation (e.g. about Percona), I have my doubts about my first conclusion.

Master and Slave DB nodes should be really consistent.
Maybe it's a problem with the Flexible Search Cache, because 1 thing is clear: the FS returns different DynamicProcessDefinitionModels depending on which Cluster Node is called. Maybe the UPD Multicast is not working as expected?