The hint of mmap usage below is really helpful to locate this problem. I forgot this biggest change for batch job in release-1.9. The blocking type option can be set to `file` as Piotr suggested to behave similar as before. I think it can solve your problem.
------------------------------------------------------------------ From:Hailu, Andreas <andreas.ha...@gs.com> Send Time:2019 Nov. 21 (Thu.) 23:37 To:Piotr Nowojski <pi...@ververica.com> Cc:Zhijiang <wangzhijiang...@aliyun.com>; user@flink.apache.org <user@flink.apache.org> Subject:RE: CoGroup SortMerger performance degradation from 1.6.4 - 1.9.1? Thanks, Piotr. We’ll rerun our apps today with this and get back to you. // ah From: Piotr Nowojski <pi...@data-artisans.com> On Behalf Of Piotr Nowojski Sent: Thursday, November 21, 2019 10:14 AM To: Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com> Cc: Zhijiang <wangzhijiang...@aliyun.com>; user@flink.apache.org Subject: Re: CoGroup SortMerger performance degradation from 1.6.4 - 1.9.1? Hi, I would suspect this: https://issues.apache.org/jira/browse/FLINK-12070 To be the source of the problems. There seems to be a hidden configuration option that avoids using memory mapped files: taskmanager.network.bounded-blocking-subpartition-type: file Could you test if helps? Piotrek On 21 Nov 2019, at 15:22, Hailu, Andreas <andreas.ha...@gs.com> wrote: Hi Zhijiang, I looked into the container logs for the failure, and didn’t see any specific OutOfMemory errors before it was killed. I ran the application using the same config this morning on 1.6.4, and it went through successfully. I took a snapshot of the memory usage from the dashboard and can send it to you if you like for reference. What stands out to me as suspicious is that on 1.9.1, the application is using nearly 6GB of Mapped memory before it dies, while 1.6.4 uses 0 throughout its runtime and succeeds. The JVM heap memory itself never exceeds its capacity, peaking at 6.65GB, so it sounds like the problem lies somewhere in the changes around mapped memory. // ah From: Zhijiang <wangzhijiang...@aliyun.com> Sent: Wednesday, November 20, 2019 11:32 PM To: Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com>; user@flink.apache.org Subject: Re: CoGroup SortMerger performance degradation from 1.6.4 - 1.9.1? Hi Andreas, You are running a batch job, so there should be no native memory used by rocked state backend. Then I guess it is either heap memory or direct memory over used. The heap managed memory is mainly used by batch operators and direct memory is used by network shuffle. Can you further check whether there are any logs to indicate HeapOutOfMemory or DirectOutOfMemory before killed? If the used memory exceeds the JVM configuration, it should throw that error. Then we can further narrow down the scope. I can not remember the changes of memory issues for managed memory or network stack, especially it really spans several releases. Best, Zhijiang ------------------------------------------------------------------ From:Hailu, Andreas <andreas.ha...@gs.com> Send Time:2019 Nov. 21 (Thu.) 01:03 To:user@flink.apache.org <user@flink.apache.org> Subject:RE: CoGroup SortMerger performance degradation from 1.6.4 - 1.9.1? Going through the release notes today - we tried fiddling with the taskmanager.memory.fraction option, going as low as 0.1 with unfortunately no success. It still leads to the container running beyond physical memory limits. // ah From: Hailu, Andreas [Engineering] Sent: Tuesday, November 19, 2019 6:01 PM To: 'user@flink.apache.org' <user@flink.apache.org> Subject: CoGroup SortMerger performance degradation from 1.6.4 - 1.9.1? Hi, We’re in the middle of testing the upgrade of our data processing flows from Flink 1.6.4 to 1.9.1. We’re seeing that flows which were running just fine on 1.6.4 now fail on 1.9.1 with the same application resources and input data size. It seems that there have been some changes around how the data is sorted prior to being fed to the CoGroup operator - this is the error that we encounter: Caused by: org.apache.flink.runtime.client.JobExecutionException: Job execution failed. at org.apache.flink.runtime.jobmaster.JobResult.toJobExecutionResult(JobResult.java:146) at org.apache.flink.client.program.rest.RestClusterClient.submitJob(RestClusterClient.java:259) ... 15 more Caused by: java.lang.Exception: The data preparation for task 'CoGroup (Dataset | Merge | NONE)' , caused an error: Error obtaining the sorted input: Thread 'SortMerger Reading Thread' terminated due to an exception: Lost connection to task manager 'd73996-213.dc.gs.com/10.47.226.218:46003'. This indicates that the remote task manager was lost. at org.apache.flink.runtime.operators.BatchTask.run(BatchTask.java:480) at org.apache.flink.runtime.operators.BatchTask.invoke(BatchTask.java:369) at org.apache.flink.runtime.taskmanager.Task.doRun(Task.java:705) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:530) ... 1 more Caused by: java.lang.RuntimeException: Error obtaining the sorted input: Thread 'SortMerger Reading Thread' terminated due to an exception: Lost connection to task manager 'd73996-213.dc.gs.com/10.47.226.218:46003'. This indicates that the remote task manager was lost. at org.apache.flink.runtime.operators.sort.UnilateralSortMerger.getIterator(UnilateralSortMerger.java:650) at org.apache.flink.runtime.operators.BatchTask.getInput(BatchTask.java:1109) at org.apache.flink.runtime.operators.CoGroupDriver.prepare(CoGroupDriver.java:102) at org.apache.flink.runtime.operators.BatchTask.run(BatchTask.java:474) I drilled further down into the YARN app logs, and I found that the container was running out of physical memory: 2019-11-19 12:49:23,068 INFO org.apache.flink.yarn.YarnResourceManager - Closing TaskExecutor connection container_e42_1574076744505_9444_01_000004 because: Container [pid=42774,containerID=container_e42_1574076744505_9444_01_000004] is running beyond physical memory limits. Current usage: 12.0 GB of 12 GB physical memory used; 13.9 GB of 25.2 GB virtual memory used. Killing container. This is what leads my suspicions as this resourcing configuration worked just fine on 1.6.4 I’m working on getting heap dumps of these applications to try and get a better understanding of what’s causing the blowup in physical memory required myself, but it would be helpful if anyone knew what relevant changes have been made between these versions or where else I could look? There are some features in 1.9 that we’d like to use in our flows so getting this sorted out, no pun intended, is inhibiting us from doing so. Best, Andreas Your Personal Data: We may collect and process information about you that may be subject to data protection laws. For more information about how we use and disclose your personal data, how we protect your information, our legal basis to use your information, your rights and who you can contact, please refer to: www.gs.com/privacy-notices Your Personal Data: We may collect and process information about you that may be subject to data protection laws. For more information about how we use and disclose your personal data, how we protect your information, our legal basis to use your information, your rights and who you can contact, please refer to: www.gs.com/privacy-notices Your Personal Data: We may collect and process information about you that may be subject to data protection laws. For more information about how we use and disclose your personal data, how we protect your information, our legal basis to use your information, your rights and who you can contact, please refer to: www.gs.com/privacy-notices
