Conversation
|
test-me-please |
|
gke seems to be failing due to unrelated cause, will restart |
|
test-gke |
2 similar comments
|
test-gke |
|
test-gke |
| // kube-proxy free case since we'll hit the wildcard rule in bpf_sock | ||
| // In-cluster connectivity from k8s2 to k8s1 IP will still work with | ||
| // HostRechableServices (regardless of if we are running with or | ||
| // without kube-proxy) since we'll hit the wildcard rule in bpf_sock |
There was a problem hiding this comment.
Could you extend the comment saying why it won't work on kube-proxy?
There was a problem hiding this comment.
Added the following comment:
// This is a known incompatibility with kube-proxy:
// kube-proxy 1.15+ will only load-balance requests from k8s1 to k8s1,
// but not from k8s2 to k8s1. In the k8s2 to k8s1 case, kube-proxy
// would send traffic to k8s1, where it would be subsequently
// dropped, because k8s1 has no service backend.
// However, if HostReachableServices is enabled, then Cilium does
// the service translation already on the client node, bypassing
// kube-proxy completely.
f9dc804 to
6c7fed3
Compare
|
test-me-please |
|
test-gke |
|
test-me-please |
`--validate` helm flag introduced in 82cc7c3 caused ci to fail in gke where we enable node init daemonset. It needs to be cleared before cilium installation. Signed-off-by: Maciej Kwiek <maciej@isovalent.com>
Signed-off-by: Sebastian Wicki <sebastian@isovalent.com>
6c7fed3 to
9575367
Compare
|
test-me-please |
|
#12725 is merged, please rebase ninja edit - didn't refresh the pr page |
|
test-gke |
|
Looks like Edit: That function probably doesn't work on GKE anways. Switched to |
This commit fixes the test harness for `externalTrafficPolicy=Local` in the case where we are accessing a node IP without a local backend from a node with host reachable services enabled. This is a known incompatibility between our kube-proxy replacement and upstream kube-proxy. The existing test harness assumed that we only need to handle this case if we are running without kube-proxy. This assumption however is wrong, as we are running these tests in hybrid mode, where we are running with both kube-proxy and Cilium's kube-proxy replacement. This has not been hit in our existing test suites up until recently, because we did not have a test setup with both kube-proxy and the kube-proxy replacement enabled at the same time. As GKE has been upgraded to Linux 4.19, it is now using the described setup which has caused the tests to break. The test matrix as of writing now looks as follows: | Test Suite | kube-proxy | kube-proxy replacement | | -------------------------- | ---------- | ----------------------- | | K8s-1.18-kernel-4.9 | Yes | No (Kernel 4.9.x) | | K8s-1.17-Kernel-4.19 | No | Yes (Kernel 4.19.57) | | K8s-1.12-Kernel-netnext | No | Yes (Kernel 5.8.0-rc1+) | | Cilium-PR-Ginkgo-Tests-k8s | Yes | No (Kernel 4.9.x) | | Cilium-PR-K8s-GKE | Yes | Yes (Kernel 4.19.112+) | Fixes: 67f85e3 ("tests: enable additional externalTrafficPolicy=Local tests") Signed-off-by: Sebastian Wicki <sebastian@isovalent.com>
9575367 to
bee0976
Compare
|
test-me-please |
|
GKE hit Suite-k8s-1.15.K8sChaosTest Connectivity demo application Endpoint can still connect while Cilium is not running which seems unrelated. |
|
test-gke |
|
Again hit Suite-k8s-1.15.K8sChaosTest Connectivity demo application Endpoint can still connect while Cilium is not running (on GKE) |
|
I'm merging this. It seems like it did fix the |
This commit fixes the test harness for
externalTrafficPolicy=Localinthe case where we are accessing a node IP without a local backend from a
node with host reachable services enabled.
This is a known incompatibility between our kube-proxy replacement and
upstream kube-proxy. The existing test harness assumed that we only need
to handle this case if we are running without kube-proxy. This
assumption however is wrong, as we are running these tests in hybrid
mode, where we are running with both kube-proxy and Cilium's kube-proxy
replacement.
This has not been hit in our existing test suites up until recently,
because we did not have a test setup with both kube-proxy and the
kube-proxy replacement enabled at the same time. As GKE has been
upgraded to Linux 4.19, it is now using the described hybrid setup which has
caused the tests to break.
The test matrix as of writing now looks as follows:
Fixes: 67f85e3 ("tests: enable additional externalTrafficPolicy=Local tests")