[ https://issues.apache.org/jira/browse/FLINK-13184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16971079#comment-16971079 ]
Yang Wang edited comment on FLINK-13184 at 11/10/19 12:24 PM: -------------------------------------------------------------- [~trohrmann] `NMClientAsync` use a internal thread pool to execute all the container launch event. When we call `startContainerAsync`, it just put a event to the blocking queue. So replacing the `NMClient` to `NMClientAsync` will not take many unknown implications. Also I will run more tests in our yarn cluster. I think using dynamic properties instead of uploading a hdfs file is better. We have some big flink applications with more than 5000 containers. Even we use a thread pool to upload the config file to hdfs, it will still be very slow. Since the config for task manager and config uploaded by flink client have a small difference, so using dynamic properties is reasonable. was (Author: fly_in_gis): [~trohrmann] `NMClientAsync` use a internal thread pool to execute all the container launch event. When we call `startContainerAsync`, it just put a event to the blocking queue. So replacing the `NMClient` to `NMClientAsync` will not take many unknown implications. Also I will run more tests in our yarn cluster. I think using dynamic properties instead of uploading a hdfs file is better. We have some big flink applications with more than 5000 containers. Even we use a thread pool to upload the config file to hdfs, it will still be very slow. Since the config for task manager and config uploaded by flink client have a small difference, so use dynamic properties is reasonable. > Starting a TaskExecutor blocks the YarnResourceManager's main thread > -------------------------------------------------------------------- > > Key: FLINK-13184 > URL: https://issues.apache.org/jira/browse/FLINK-13184 > Project: Flink > Issue Type: Bug > Components: Deployment / YARN > Affects Versions: 1.8.1, 1.9.0, 1.10.0 > Reporter: Xintong Song > Priority: Critical > Labels: pull-request-available > Fix For: 1.10.0, 1.8.3, 1.9.2 > > Time Spent: 0.5h > Remaining Estimate: 0h > > Currently, YarnResourceManager starts all task executors in main thread. This > could cause RM to become unresponsive when launching a large number of TEs > (e.g. > 1000) because it involves blocking I/O operations (writing files to > HDFS, communicating with the node manager using a synchronous {{NMClient}}). > As a consequence, TE registration/heartbeat timeouts can occur and Flink > might allocate too many excessive containers (see FLINK-12342) because it > cannot process the {{YarnResourceManager#onContainersAllocated}} calls. > There are different solution approaches but the end goal should be to not > execute any blocking calls in the {{ResourceManager's}} main thread: > 1. Start the TaskExecutors from a different thread (potentially thread pool) > which is responsible for uploading the files and communicating with the > NodeManager > 2. Don't upload files (avoid blocking file system operations) and use the > {{NMClientAsync}} for the communication with Yarn's {{NodeManager}}. > 3. Upload files in a separate I/O thread and use the {{NMClientAsync}} for > the communication with Yarn's {{NodeManager}}. -- This message was sent by Atlassian Jira (v8.3.4#803005)