yarn 单点故障 重启 ResourceManger Restart

Posted 2021-12-14 yuanjiangw

 

http://hadoop.apache.org/docs/stable/hadoop-yarn/hadoop-yarn-site/ResourceManagerRestart.html

 

Feature

• Phase 1: Non-work-preserving RM restart

  As of Hadoop 2.4.0 release, only ResourceManager Restart Phase 1 is implemented which is described below.

  The overall concept is that RM will persist the application metadata (i.e. ApplicationSubmissionContext) in a pluggable state-store when client submits an application and also saves the final status of the application such as the completion state (failed, killed, finished) and diagnostics when the application completes. Besides, RM also saves the credentials like security keys, tokens to work in a secure environment. Any time RM shuts down, as long as the required information (i.e.application metadata and the alongside credentials if running in a secure environment) is available in the state-store, when RM restarts, it can pick up the application metadata from the state-store and re-submit the application. RM won’t re-submit the applications if they were already completed (i.e. failed, killed, finished) before RM went down.

  NodeManagers and clients during the down-time of RM will keep polling RM until RM comes up. When RM becomes alive, it will send a re-sync command to all the NodeManagers and ApplicationMasters it was talking to via heartbeats. As of Hadoop 2.4.0 release, the behaviors for NodeManagers and ApplicationMasters to handle this command are: NMs will kill all its managed containers and re-register with RM. From the RM’s perspective, these re-registered NodeManagers are similar to the newly joining NMs. AMs(e.g. MapReduce AM) are expected to shutdown when they receive the re-sync command. After RM restarts and loads all the application metadata, credentials from state-store and populates them into memory, it will create a new attempt (i.e. ApplicationMaster) for each application that was not yet completed and re-kick that application as usual. As described before, the previously running applications’ work is lost in this manner since they are essentially killed by RM via the re-sync command on restart.

• Phase 2: Work-preserving RM restart

  As of Hadoop 2.6.0, we further enhanced RM restart feature to address the problem to not kill any applications running on YARN cluster if RM restarts.

  Beyond all the groundwork that has been done in Phase 1 to ensure the persistency of application state and reload that state on recovery, Phase 2 primarily focuses on re-constructing the entire running state of YARN cluster, the majority of which is the state of the central scheduler inside RM which keeps track of all containers’ life-cycle, applications’ headroom and resource requests, queues’ resource usage etc. In this way, RM doesn’t need to kill the AM and re-run the application from scratch as it is done in Phase 1. Applications can simply re-sync back with RM and resume from where it were left off.

  RM recovers its runing state by taking advantage of the container statuses sent from all NMs. NM will not kill the containers when it re-syncs with the restarted RM. It continues managing the containers and send the container statuses across to RM when it re-registers. RM reconstructs the container instances and the associated applications’ scheduling status by absorbing these containers’ information. In the meantime, AM needs to re-send the outstanding resource requests to RM because RM may lose the unfulfilled requests when it shuts down. Application writers using AMRMClient library to communicate with RM do not need to worry about the part of AM re-sending resource requests to RM on re-sync, as it is automatically taken care by the library itself.