I'm dealing with a very high-volume workload where a Premium Event Hubs namespace with 16 PUs and 100 partitions may not provide enough capacity. Would it be possible to place an Azure Load Balancer in front of three Premium Event Hubs namespaces and distribute producer traffic across them? The clients use TCP connections on port 9093, so I'm also wondering whether that creates any limitations. Are there ingress throughput limits on Azure Load Balancer that would affect this design?
2 Answers
An Azure Load Balancer probably isn’t suitable for this architecture. TCP port 9093 isn’t the main issue; Event Hubs namespaces or their private endpoints can’t be used as backend pool targets for an Azure Load Balancer. Azure Load Balancer is primarily intended for IaaS resources such as virtual machines and virtual machine scale sets.
For a workload at this scale, Event Hubs Dedicated is worth evaluating because it is designed for higher and more predictable throughput. Another option is to deploy multiple Premium namespaces and explicitly shard producers across them, but that moves the routing and balancing logic into your application. I’d benchmark that approach against Dedicated before making a decision.
Port 9093 itself shouldn’t prevent a TCP load-balancing design, but the backend compatibility is the limiting factor here. An Azure Load Balancer can distribute traffic to supported compute resources, not directly across Event Hubs namespaces. If you use multiple Premium namespaces, your producers will need to select a namespace through application-level routing or a separate proxy layer. Be sure to test partitioning, ordering, retries, and consumer-group management across namespaces, since those concerns become your responsibility.
That matches my concern. Dedicated seems substantially more expensive than Premium, though, and I’m still unclear about how Dedicated Capacity Units compare with Premium Throughput Units. Is there a practical CU-to-PU conversion, or does the capacity depend too much on the workload for a direct comparison? We may instead assign different producers and consumer groups to separate Premium namespaces and handle the sharding ourselves.