control-plane: implement CDS and EDS for universal case #218
Conversation
yskopets
changed the title
| } | ||
| for _, dataplane := range dataplanes.Items { | ||
| for _, inbound := range dataplane.Spec.Networking.GetInbound() { | ||
| service := inbound.Tags[mesh_proto.ServiceTag] |
There was a problem hiding this comment.
We should match also other tags.
| return nil, err | ||
| } | ||
|
|
||
| edsClusterName := fmt.Sprintf("%s:%d", endpoint.DataplaneIP, endpoint.DataplanePort) |
There was a problem hiding this comment.
Cluster name should be service name. Otherwise it will be really hard to read metrics later. It's common to do an aggregation by the service name from all Envoys.
| @@ -66,9 +68,16 @@ func DefaultDataplaneSyncTracker(rt core_runtime.Runtime, reconciler SnapshotRec | |||
| }, | |||
| } | |||
|
|
|||
| outbound, err := xds_topology.GetOutboundTargets(ctx, dataplane, rt.ResourceManager()) | |||
There was a problem hiding this comment.
I see we do it for every dataplane every X seconds. This would be a bottleneck very soon (even with caching on resource store).
What I imagine is that we've got a job (Component=Runnable) that builds Cluster and ClusterLoadAssignment from Dataplanes and expose it to the OutboundProxyGenerator.
It's more efficient in CPU, because we do the matching once (every X second) and then OutboundProxyGenerator can just retrieve this in O(1).
It's more efficient in memory. Every Snapshot will be stored in SnapshotCache. I don't know about Go but when I introduced reusing protos in java-control-plane based control plane, there performance went through the roof
envoyproxy/java-control-plane#98
changes:
universalcase