Clayton Coleman
Clayton Coleman
https://github.com/kubernetes/kubernetes/pull/98507#issuecomment-845197777 connects the two issues - that PR doesn't go far enough because there are bugs in other loops (loops are confused between "fully terminated" and "terminating", and the pod...
> The scenario described here was discussed and warned at SIG Node several times long time ago, and one particular example I were using is disk space clean up after...
Now that this is getting close to complete, moving the pod worker part to #102344 - please take all review there.
This PR is *fixing* the bugs in the e2e tests. It was the fact that we were abusing this in tests, and then that a large set of flaky issues...
Talking with jordan he suggested three things: 1. a warning 2. a feature gate to control the change 3. a metric all of which are reasonable as we are changing...
> Unfortunately this would probably also break most or all clients parsing it, should we actually use more than 63 bits. Ideally, if we wanted to "reclaim resourceVersion" we'd probably...
> Today you can accomplish this by incrementing the etcd global index prior to reconnecting the cluster Right, if you know the increment of your lost control plane, which in...