When running nodepool launchers in kubernetes a common method to update nodepool or its config is doing rolling restarts. The process for this is start a new nodepool, wait for it to be ready and then tear down the old instance. Currently this is not possible without risking node_failures when there is only one instance serving a label. The reason for this is that there is no reliable way to determine when the new instance is fully started which could lead to a too early tear down of the old instance. This would result in node_failures for all in-flight nore requests that are only valid for this provider. Adding a /ready endpoint to the webapp can make this deterministic using readiness checks of kubernetes. Change-Id: I53e77f3d8aaa4742ce2a89c1179e8563f850270e
110 B
110 B