|
Spyros, Do you have a link handy to the etcd bug? Thx, britt From:
<atomic-devel-bounces projectatomic io> on behalf of Spyros Trigazis <strigazi gmail com> Hi all. Recently, I discovered a bug on etcd which had a serious performance impact. When etcd 3.0.x was built with golang 1.7.x, it was throwing the following exception constantly: etcd[2386]: transport: http2Client.notifyError got notified that the client transport was broken unexpected EOF.
This bug was introduced because of golang v1.7.x and was fixed in etcd v3.1.x. In our kubernetes cluster configuration, we have some master nodes which contain, etcd, kube-apiserver, kube-scheduler and kube-controller-manager. Because of the above bug, the apiserver wasn't responding due to high cpu load coming from etcd. So, in any application, if etcd is running in powerless machine, will be unresponsive or if other applications run in the same machine their performance will be affected. I propose to add a test for etcd, configured with TLS certs similar to my reproducer for this bug [1]. The reasoning is that, this bug appeared only when etcd was configured with TLS, other bugs might occur with that setting in the future. I'm new to the packaging process of fedora and I don't know where to start. Cheers, Spyros |