We’re planning to deploy a service to our mesh that will serve 1 million+ concurrent, persistent SSE connections. During initial load testing, our sidecars died with out-of-memory errors after handling around 100 connections each under the default sidecar resource allocation and pretty light load.
While we can increase our sidecar resources with annotations (thanks for https://github.com/istio/istio/pull/22468), we’re wondering if it might be better to deploy this particular service outside the mesh. Has anybody deployed a similar service in their mesh with a high number of persistent SSE connections, or even persistent TCP connections? Very interested in your thoughts, thanks.