Hi Xintong and Stephan, Thanks a lot for your attention on this FLIP. I will address the comments inline.
# Architecture -> One or two ConfigMaps Both of you are right. One ConfigMap will make the design and implementation easier. Actually, in my POC codes, I am using just one ConfigMap(e.g. "k8s-ha-app1-restserver" for rest server component) for the leader election and storage. Once a JobManager win the election, it will update the ConfigMap with leader address and periodically renew the lock annotation to keep as the active leader. I will update the FLIP document, including the architecture diagram, to avoid the misunderstanding. # HA storage > Lock and release This is a valid concern. Since for Zookeeper ephemeral nodes, it will be deleted by the ZK server automatically when the client is timeout. It could happen in a bad network environment or the ZK client crashed exceptionally. For Kubernetes, we need to implement a similar mechanism. First, when we want to lock a specific key in ConfigMap, we will put the owner identify, lease duration, renew time in the ConfigMap annotation. The annotation will be cleaned up when releasing the lock. When we want to remove a job graph or checkpoints, it should satisfy the following conditions. If not, the delete operation could not be done. * Current instance is the owner of the key. * The owner annotation is empty, which means the owner has released the lock. * The owner annotation timed out, which usually indicate the owner died. # HA storage > HA data clean up Sorry for that I do not describe how the HA related ConfigMap is retained clearly. Benefit from the Kubernetes OwnerReference[1], we set owner of the flink-conf configmap, service and TaskManager pods to JobManager Deployment. So when we want to destroy a Flink cluster, we just need to delete the deployment[2]. For the HA related ConfigMaps, we do not set the owner so that they could be retained even though we delete the whole Flink cluster. [1]. https://kubernetes.io/docs/concepts/workloads/controllers/garbage-collection/ [2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html#stop-flink-session Best, Yang Stephan Ewen <se...@apache.org> 于2020年9月16日周三 下午8:16写道： > This is a very cool feature proposal. > > One lesson-learned from the ZooKeeper-based HA is that it is overly > complicated to have the Leader RPC address in a different node than the > LeaderLock. There is extra code needed to make sure these converge and the > can be temporarily out of sync. > > A much easier design would be to have the RPC address as payload in the > lock entry (ZNode in ZK), the same way that the leader fencing token is > stored as payload of the lock. > I think for the design above it would mean having a single ConfigMap for > both leader lock and leader RPC address discovery. > > This probably serves as a good design principle in general - not divide > information that is updated together over different resources. > > Best, > Stephan > > > On Wed, Sep 16, 2020 at 11:26 AM Xintong Song <tonysong...@gmail.com> > wrote: > >> Thanks for preparing this FLIP, @Yang. >> >> In general, I'm +1 for this new feature. Leveraging Kubernetes's buildtin >> ConfigMap for Flink's HA services should significantly reduce the >> maintenance overhead compared to deploying a ZK cluster. I think this is an >> attractive feature for users. >> >> Concerning the proposed design, I have some questions. Might not be >> problems, just trying to understand. >> >> ## Architecture >> >> Why does the leader election need two ConfigMaps (`lock for contending >> leader`, and `leader RPC address`)? What happens if the two ConfigMaps are >> not updated consistently? E.g., a TM learns about a new JM becoming leader >> (lock for contending leader updated), but still gets the old leader's >> address when trying to read `leader RPC address`? >> >> ## HA storage > Lock and release >> >> It seems to me that the owner needs to explicitly release the lock so >> that other peers can write/remove the stored object. What if the previous >> owner failed to release the lock (e.g., dead before releasing)? Would there >> be any problem? >> >> ## HA storage > HA data clean up >> >> If the ConfigMap is destroyed on `kubectl delete deploy <ClusterID>`, how >> are the HA dada retained? >> >> >> Thank you~ >> >> Xintong Song >> >> >> >> On Tue, Sep 15, 2020 at 11:26 AM Yang Wang <danrtsey...@gmail.com> wrote: >> >>> Hi devs and users, >>> >>> I would like to start the discussion about FLIP-144[1], which will >>> introduce >>> a new native high availability service for Kubernetes. >>> >>> Currently, Flink has provided Zookeeper HA service and been widely used >>> in production environments. It could be integrated in standalone cluster, >>> Yarn, Kubernetes deployments. However, using the Zookeeper HA in K8s >>> will take additional cost since we need to manage a Zookeeper cluster. >>> In the meantime, K8s has provided some public API for leader election[2] >>> and configuration storage(i.e. ConfigMap[3]). We could leverage these >>> features and make running HA configured Flink cluster on K8s more >>> convenient. >>> >>> Both the standalone on K8s and native K8s could benefit from the new >>> introduced KubernetesHaService. >>> >>> [1]. >>> https://cwiki.apache.org/confluence/display/FLINK/FLIP-144%3A+Native+Kubernetes+HA+for+Flink >>> [2]. >>> https://kubernetes.io/blog/2016/01/simple-leader-election-with-kubernetes/ >>> [3]. https://kubernetes.io/docs/concepts/configuration/configmap/ >>> >>> Looking forward to your feedback. >>> >>> Best, >>> Yang >>> >>
