You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Chart 2.0 ships per-component Ingress (dashboard.ingress, backend.split.management.ingress.{http,grpc}, backend.split.signal.ingress, backend.split.relay.ingress) but nothing for Gateway API. Clusters on Traefik, Envoy Gateway, Cilium or Istio Gateway currently have to hand-write HTTPRoutes next to the release, and ingress-nginx retirement (#27) pushes more installs that way.
#51 added routes for the 1.x layout but conflicts since the rewrite in #54.
Proposal: an httpRoute block per component, mirroring the existing ingress blocks:
dashboard.httpRoute
backend.split.management.httpRoute.http (/api, /oauth2) and .grpc (/management.ManagementService, /management.ProxyService)
each with enabled, parentRefs, hostnames, annotations, and the chart's default paths.
Open question: management REST and gRPC share one Service port that carries appProtocol: kubernetes.io/h2c (#63), so a gateway also sends /api to it over h2c. Is REST over h2c supported by management, or should the chart offer a separate gRPC Service port for the gRPC route?
On the open question — I dug into where the h2c marking comes from, and I think it decides the shape of the whole httpRoute block, so it's worth settling before anyone writes templates.
Today appProtocol: kubernetes.io/h2c is set on exactly two ports (render.sh check 17 pins the count at 2): management http and signal grpc. For management that one port serves both the REST API and the gRPC services, so a gateway that honours appProtocol will open HTTP/2 prior-knowledge connections for /api and /oauth2 as well, not just for /management.ManagementService.
That leaves three ways to go, and they produce different values:
One route, one port — if management's REST handler is served by the same HTTP/2 server, h2c for /api is fine and management.httpRoute can stay a single block with both path sets.
Split the Service port — add a separate gRPC port so REST keeps HTTP/1.1 and only the gRPC port carries h2c. This changes the Service for Ingress users too, not just Gateway API, so it's a bigger call than the route block itself.
Keep one port, make appProtocol per-route — leave the Service alone and let the route decide, which works on some implementations and not others.
Could a maintainer confirm whether management's REST API is served over h2c in practice? If it is, (1) keeps this to a pure addition with no change for existing users, which seems like the right default.
Happy to implement whichever way you land on. I've just sent #75, #76 and #77 for #73, #21 and #19, so I have the 2.0 layout loaded — but I didn't want to guess at this one and send a large set of templates built on the wrong assumption.
Activity
codechirag5610 commentedon Nov 16, 2026
On the open question — I dug into where the h2c marking comes from, and I think it decides the shape of the whole
httpRouteblock, so it's worth settling before anyone writes templates.Today
appProtocol: kubernetes.io/h2cis set on exactly two ports (render.shcheck 17 pins the count at 2): managementhttpand signalgrpc. For management that one port serves both the REST API and the gRPC services, so a gateway that honoursappProtocolwill open HTTP/2 prior-knowledge connections for/apiand/oauth2as well, not just for/management.ManagementService.That leaves three ways to go, and they produce different values:
/apiis fine andmanagement.httpRoutecan stay a single block with both path sets.appProtocolper-route — leave the Service alone and let the route decide, which works on some implementations and not others.Could a maintainer confirm whether management's REST API is served over h2c in practice? If it is, (1) keeps this to a pure addition with no change for existing users, which seems like the right default.
Happy to implement whichever way you land on. I've just sent #75, #76 and #77 for #73, #21 and #19, so I have the 2.0 layout loaded — but I didn't want to guess at this one and send a large set of templates built on the wrong assumption.