Every time we added a model lane behind our gateway, the bottleneck wasn't only GPUs. It was certificates too.
Each lane had to be distinguished by hostname or port โ new SAN entry, DNS record, cert rotation overhead. On managed cert infra, that's the tax on adding a model.
In multi-cluster llm-d, suppose the EPP on cluster A steers requests to gateways on clusters B and C. Those gateways have DNS names. But the proxy only understood IP:port. So we had to hardcode load
balancer IPs โ which break silently when an LB is recreated.
Open source contributions is more about convincing maintainers for the issue and approach, rather than implementation, felt this most of the time. ๐ซก
#opensource
@MohitMa11446106 yes, even more secure as it comes under the L7 TLS protect communication which can be served under VPN persistently. It will be somewhat slow(few milli seconds), but that gap we can ignore.