withdraw ClusterIP bgp route for externalTrafficPolicy=Local when there are no local endpoints #347
Conversation
|
Refs #347 |
| err := nrc.bgpServer.DeletePath([]byte(nil), 0, "", pathList) | ||
|
|
||
| if err != nil { | ||
| return fmt.Errorf(err.Error()) |
There was a problem hiding this comment.
I agree with you, but the AdvertiseClusterIp method above it should be refactored as well then, since the methods are structured the same
There was a problem hiding this comment.
feel free to fix the method above too :)
|
Tested this patch and it works as we expect. However, the convergence time in kube-router is taking a bit longer than I would like once the route is removed. I think this is expected if you use the default sync period (60s) |
|
Might be a good time to refactor to use informers so we can get the update events for endpoints right away, as opposed to listing all endpoints on each sync loop |
|
I agree that the convergence is not really acceptable. Ideally it should be ran immediately when there are no endpoints left on the node. |
|
LGTM, thanks for the PR @TvL2386 I'll try to do some work in the upcoming week to shorten the convergence time. |
Hi guys,
I noticed that when using a Service object with type: Loadbalancer and externalTrafficPolicy: Local, the routes are not withdrawn. On my external BGP peer I could still see the route which caused for dropped packets when that route was used.
This pull request withdraws routes for this externalTrafficPolicy when there are no endpoints.