Event ordering in Kafka: partition-level ordering vs global ordering

we're running into complexities with event ordering in kafka. for most of our use cases, partitioning by `userId` gives us good enough ordering guarantees within a partition, which is fine for user-specific event streams. however, we have some events, like an admin disabling an account, where global ordering across all users might be implicitly expected by downstream consumers or reporting systems. forcing everything into a single partition for global ordering is obviously not scalable for throughput. are there patterns people use to handle occasional global ordering requirements without sacrificing parallelism? perhaps a separate topic for globally ordered events, or some form of re-partitioning downstream that doesn't introduce excessive latency or complexity? what are the risks and trade-offs when trying to achieve a semblance of global order in a distributed event stream?