Hi,
we had a similar issue. Multi-instance with Camunda does not work well when it comes to thousands of instances even if the task has no variables and just prints “Hello World”. See Scaling multi-instance tasks, true parallel-execution and play around with https://github.com/marsewe/camunda-external-task-test .
Scaling with Camunda though works from my experience when you scale via the number of process instances, a 100k running instances work smoothly.