Hi all,
We would like to open a discussion on merging the Router-based Federation feature to trunk. Last week, there was a thread about which branches would go into 3.0 and given that YARN federation is going, this might be a good time for this to be merged too. We have been running "Router-based federation" in production for a year. Meanwhile, we have been releasing it in a feature branch (HDFS-10467 [1]) for a while. We are reasonably confident that the state of the branch is about to meet the criteria to be merged onto trunk. *Feature*: This feature aggregates multiple namespaces into a single one transparently to the user. It has a similar architecture to YARN federation (YARN-2915). It consists on Routers that handle requests from the clients and forwards them to the right subcluster and exposes the same API as the Namenode. Currently we use a mount table (similar to ViewFs) but can be replaced by other approaches. The Routers share their state in a State Store. The main advantage is that clients interact with the Routers as they were Namenode so there is no changes in the client required other than poiting to the right address. In addition, all the management is moved to the server side so changes to the Mount Table can be done without having to sync the clients (pull/push). *Status*: The branch already contains all the features required to work end-to-end. There are a couple open JIRAs that would be required for the merged (i.e., Web UI) but they should be finished soon. We have been running it in production for the last year and we have a paper with some of the details of our production deployment [2]. We have 4 production deployments with the largest one spanning more than 20k servers across 6 subclusters. In addition, the guys at LinkedIn had started testing Router-based federation and they will be adding security to the branch. The modifications to the rest of HDFS are minimal: - Changed visibility for some methods (e.g., MiniDFSCluster) - Added some utilities to extract addresses - Modified hdfs and hdfs.cmd to start the Router and manager the federation - Modified hdfs-default.xml Everything else is self-contained in a federation package. In addition, all the functionality is in the Router so it’s disabled by default. Even when enabled, there is no impact for regular HDFS and it would only require to configure the trust between the Namenode and the Router once security is enabled. I have been continuously rebasing the feature branch (updated up to 1 week ago) so the merge should be a straightforward cherry-pick. *Problems*: The problems I’m aware of are the following: - We implement ClientProtocol so anytime a new method is added there, we would need to add it to the Router. However, it’s straightforward to add unimplemented methods. - There is some argument about naming the feature as “Router-based federation” but I’m open for better names. *Credits*: I’d like to thank the people at Microsoft (specially, Jason, Ricardo, Chris, Subru, Jakob, Carlo and Giovanni), Twitter (Ming and Gera), and LinkedIn (Zhe, Erik and Konstantin) for the discussion and the ideas. Special thanks to Chris Douglas for the thorough reviews! Please look through the branch; feedback is welcome. Thanks! Cheers, Inigo [1] https://issues.apache.org/jira/browse/HDFS-10467 [2] https://www.usenix.org/conference/atc17/technical- sessions/presentation/misra
