Hi Mark, Have you tried to set your rolling policy to close inactive part files after some time [1]? If the part files in the buckets are inactive and there are no new part files, then the state handle for those buckets will also be removed.
Cheers, Kostas https://ci.apache.org/projects/flink/flink-docs-master/api/java/org/apache/flink/streaming/api/functions/sink/filesystem/rollingpolicies/DefaultRollingPolicy.html On Mon, Feb 3, 2020 at 3:54 PM Mark Harris <mark.har...@hivehome.com> wrote: > Hi all, > > The out-of-memory heap dump had the answer - the job was failing with an > OutOfMemoryError because the activeBuckets members of 3 instances of > org.apache.flink.streaming.api.functions.sink.filesystem.Buckets were > filling a significant enough part of the memory of the taskmanager that no > progress was being made. Increasing the memory available to the TM seems to > have fixed the problem. > > I think the DeleteOnExit problem will mean it needs to be restarted every > few weeks, but that's acceptable for now. > > Thanks again, > > Mark > ------------------------------ > *From:* Mark Harris <mark.har...@hivehome.com> > *Sent:* 30 January 2020 14:36 > *To:* Piotr Nowojski <pi...@ververica.com> > *Cc:* Cliff Resnick <cre...@gmail.com>; David Magalhães < > speeddra...@gmail.com>; Till Rohrmann <trohrm...@apache.org>; > flink-u...@apache.org <flink-u...@apache.org>; kkloudas < > kklou...@apache.org> > *Subject:* Re: GC overhead limit exceeded, memory full of DeleteOnExit > hooks for S3a files > > Hi, > > Thanks for your help with this. 🙂 > > The EMR cluster has 3 15GB VMs, and the flink cluster is started with: > > /usr/lib/flink/bin/yarn-session.sh -d -n 3 -tm 5760 -jm 5760 -s 3 > > Usually the task runs for about 15 minutes before it restarts, usually due > to with an "java.lang.OutOfMemoryError: Java heap space" exception. > > The figures came from a MemoryAnalyzer session on a manual memory dump > from one of the taskmanagers. The total size of that heap was only 1.8gb. > In that heap, 1.7gb is taken up by the static field "files" in > DeleteOnExitHook, which is a linked hash set containing the 9 million > strings. > > A full example of one the path is > /tmp/hadoop-yarn/s3a/s3ablock-0001-6061210725685.tmp, at for 120 bytes per > char[] for a solid 1.2gb of chars. Then 200mb for their String wrappers and > another 361MB for LinkedHashMap$Entry objects. Despite valiantly holding > on to an array of 16777216 HashMap$Node elements, the LinkedHashMap can > only contribute another 20MB or so. > I goofed in not taking that 85% figure from MemoryAnalyzer - it tells > me DeleteOnExitHook is responsible for 96.98% of the heap dump. > > Looking at the files it managed to write before this started to happen > regularly, it looks like they're being written approximately every 3 > minutes. I'll triple check our config, but I'm reasonably sure the job is > configured to checkpoint every 15 minutes - could something else be causing > it to write? > > This may all be a red herring - something else may be taking up the > taskmanagers memory which didn't make it into that heap dump. I plan to > repeat the analysis on a heapdump created > by -XX:+HeapDumpOnOutOfMemoryError shortly. > > Best regards, > > Mark > > ------------------------------ > *From:* Piotr Nowojski <pi...@ververica.com> > *Sent:* 30 January 2020 13:44 > *To:* Mark Harris <mark.har...@hivehome.com> > *Cc:* Cliff Resnick <cre...@gmail.com>; David Magalhães < > speeddra...@gmail.com>; Till Rohrmann <trohrm...@apache.org>; > flink-u...@apache.org <flink-u...@apache.org>; kkloudas < > kklou...@apache.org> > *Subject:* Re: GC overhead limit exceeded, memory full of DeleteOnExit > hooks for S3a files > > Hi, > > What is your job setup? Size of the nodes, memory settings of the > Flink/JVM? > > 9 041 060 strings is awfully small number to bring down a whole cluster. > With each tmp string having ~30 bytes, that’s only 271MB. Is this really > 85% of the heap? And also, with parallelism of 6 and checkpoints every 15 > minutes, 9 000 000 of leaked strings should happen only after one month > assuming 500-600 total number of buckets. (Also assuming that there is a > separate file per each bucket). > > Piotrek > > On 30 Jan 2020, at 14:21, Mark Harris <mark.har...@hivehome.com> wrote: > > Trying a few different approaches to the fs.s3a.fast.upload settings has > bought me no joy - the taskmanagers end up simply crashing or complaining > of high GC load. Heap dumps suggest that this time they're clogged with > buffers instead, which makes sense. > > Our job has parallelism of 6 and checkpoints every 15 minutes - if > anything, we'd like to increase the frequency of that checkpoint duration. > I suspect this could be affected by the partition structure we were > bucketing to as well, and at any given moment we could be receiving data > for up to 280 buckets at once. > Could this be a factor? > > Best regards, > > Mark > ------------------------------ > *From:* Piotr Nowojski <pi...@ververica.com> > *Sent:* 27 January 2020 16:16 > *To:* Cliff Resnick <cre...@gmail.com> > *Cc:* David Magalhães <speeddra...@gmail.com>; Mark Harris < > mark.har...@hivehome.com>; Till Rohrmann <trohrm...@apache.org>; > flink-u...@apache.org <flink-u...@apache.org>; kkloudas < > kklou...@apache.org> > *Subject:* Re: GC overhead limit exceeded, memory full of DeleteOnExit > hooks for S3a files > > Hi, > > I think reducing the frequency of the checkpoints and decreasing > parallelism of the things using the S3AOutputStream class, would help to > mitigate the issue. > > I don’t know about other solutions. I would suggest to ask this question > directly to Steve L. in the bug ticket [1], as he is the one that fixed the > issue. If there is no workaround, maybe it would be possible to put a > pressure on the Hadoop guys to back port the fix to older versions? > > Piotrek > > [1] https://issues.apache.org/jira/browse/HADOOP-15658 > > On 27 Jan 2020, at 15:41, Cliff Resnick <cre...@gmail.com> wrote: > > I know from experience that Flink's shaded S3A FileSystem does not > reference core-site.xml, though I don't remember offhand what file (s) it > does reference. However since it's shaded, maybe this could be fixed by > building a Flink FS referencing 3.3.0? Last I checked I think it referenced > 3.1.0. > > On Mon, Jan 27, 2020, 8:48 AM David Magalhães <speeddra...@gmail.com> > wrote: > > Does StreamingFileSink use core-site.xml ? When I was using it, it didn't > load any configurations from core-site.xml. > > On Mon, Jan 27, 2020 at 12:08 PM Mark Harris <mark.har...@hivehome.com> > wrote: > > Hi Piotr, > > Thanks for the link to the issue. > > Do you know if there's a workaround? I've tried setting the following in > my core-site.xml: > > fs.s3a.fast.upload.buffer=true > > To try and avoid writing the buffer files, but the taskmanager breaks with > the same problem. > > Best regards, > > Mark > ------------------------------ > *From:* Piotr Nowojski <pi...@data-artisans.com> on behalf of Piotr > Nowojski <pi...@ververica.com> > *Sent:* 22 January 2020 13:29 > *To:* Till Rohrmann <trohrm...@apache.org> > *Cc:* Mark Harris <mark.har...@hivehome.com>; flink-u...@apache.org < > flink-u...@apache.org>; kkloudas <kklou...@apache.org> > *Subject:* Re: GC overhead limit exceeded, memory full of DeleteOnExit > hooks for S3a files > > Hi, > > This is probably a known issue of Hadoop [1]. Unfortunately it was only > fixed in 3.3.0. > > Piotrek > > [1] https://issues.apache.org/jira/browse/HADOOP-15658 > > On 22 Jan 2020, at 13:56, Till Rohrmann <trohrm...@apache.org> wrote: > > Thanks for reporting this issue Mark. I'm pulling Klou into this > conversation who knows more about the StreamingFileSink. @Klou does the > StreamingFileSink relies on DeleteOnExitHooks to clean up files? > > Cheers, > Till > > On Tue, Jan 21, 2020 at 3:38 PM Mark Harris <mark.har...@hivehome.com> > wrote: > > Hi, > > We're using flink 1.7.2 on an EMR cluster v emr-5.22.0, which runs hadoop > v "Amazon 2.8.5". We've recently noticed that some TaskManagers fail > (causing all the jobs running on them to fail) with an > "java.lang.OutOfMemoryError: GC overhead limit exceeded”. The taskmanager > (and jobs that should be running on it) remain down until manually > restarted. > > I managed to take and analyze a memory dump from one of the afflicted > taskmanagers. > > It showed that 85% of the heap was made up of > the java.io.DeleteOnExitHook.files hashset. The majority of the strings in > that hashset (9041060 out of ~9041100) pointed to files that began > /tmp/hadoop-yarn/s3a/s3ablock > > The problem seems to affect jobs that make use of the StreamingFileSink - > all of the taskmanager crashes have been on the taskmaster running at least > one job using this sink, and a cluster running only a single taskmanager / > job that uses the StreamingFileSink crashed with the GC overhead limit > exceeded error. > > I've had a look for advice on handling this error more broadly without > luck. > > Any suggestions or advice gratefully received. > > Best regards, > > Mark Harris > > > > The information contained in or attached to this email is intended only > for the use of the individual or entity to which it is addressed. If you > are not the intended recipient, or a person responsible for delivering it > to the intended recipient, you are not authorised to and must not disclose, > copy, distribute, or retain this message or any part of it. It may contain > information which is confidential and/or covered by legal professional or > other privilege under applicable law. > > The views expressed in this email are not necessarily the views of > Centrica plc or its subsidiaries, and the company, its directors, officers > or employees make no representation or accept any liability for its > accuracy or completeness unless expressly stated to the contrary. > > Additional regulatory disclosures may be found here: > https://www.centrica.com/privacy-cookies-and-legal-disclaimer#email > > PH Jones is a trading name of British Gas Social Housing Limited. British > Gas Social Housing Limited (company no: 01026007), British Gas Trading > Limited (company no: 03078711), British Gas Services Limited (company no: > 3141243), British Gas Insurance Limited (company no: 06608316), British Gas > New Heating Limited (company no: 06723244), British Gas Services > (Commercial) Limited (company no: 07385984) and Centrica Energy (Trading) > Limited (company no: 02877397) are all wholly owned subsidiaries of > Centrica plc (company no: 3033654). Each company is registered in England > and Wales with a registered office at Millstream, Maidenhead Road, Windsor, > Berkshire SL4 5GD. > > British Gas Insurance Limited is authorised by the Prudential Regulation > Authority and regulated by the Financial Conduct Authority and the > Prudential Regulation Authority. British Gas Services Limited and Centrica > Energy (Trading) Limited are authorised and regulated by the Financial > Conduct Authority. British Gas Trading Limited is an appointed > representative of British Gas Services Limited which is authorised and > regulated by the Financial Conduct Authority. > > > > > The information contained in or attached to this email is intended only > for the use of the individual or entity to which it is addressed. If you > are not the intended recipient, or a person responsible for delivering it > to the intended recipient, you are not authorised to and must not disclose, > copy, distribute, or retain this message or any part of it. It may contain > information which is confidential and/or covered by legal professional or > other privilege under applicable law. > > The views expressed in this email are not necessarily the views of > Centrica plc or its subsidiaries, and the company, its directors, officers > or employees make no representation or accept any liability for its > accuracy or completeness unless expressly stated to the contrary. > > Additional regulatory disclosures may be found here: > https://www.centrica.com/privacy-cookies-and-legal-disclaimer#email > > PH Jones is a trading name of British Gas Social Housing Limited. British > Gas Social Housing Limited (company no: 01026007), British Gas Trading > Limited (company no: 03078711), British Gas Services Limited (company no: > 3141243), British Gas Insurance Limited (company no: 06608316), British Gas > New Heating Limited (company no: 06723244), British Gas Services > (Commercial) Limited (company no: 07385984) and Centrica Energy (Trading) > Limited (company no: 02877397) are all wholly owned subsidiaries of > Centrica plc (company no: 3033654). Each company is registered in England > and Wales with a registered office at Millstream, Maidenhead Road, Windsor, > Berkshire SL4 5GD. > > British Gas Insurance Limited is authorised by the Prudential Regulation > Authority and regulated by the Financial Conduct Authority and the > Prudential Regulation Authority. British Gas Services Limited and Centrica > Energy (Trading) Limited are authorised and regulated by the Financial > Conduct Authority. British Gas Trading Limited is an appointed > representative of British Gas Services Limited which is authorised and > regulated by the Financial Conduct Authority. > > > > > The information contained in or attached to this email is intended only > for the use of the individual or entity to which it is addressed. If you > are not the intended recipient, or a person responsible for delivering it > to the intended recipient, you are not authorised to and must not disclose, > copy, distribute, or retain this message or any part of it. It may contain > information which is confidential and/or covered by legal professional or > other privilege under applicable law. > > The views expressed in this email are not necessarily the views of > Centrica plc or its subsidiaries, and the company, its directors, officers > or employees make no representation or accept any liability for its > accuracy or completeness unless expressly stated to the contrary. > > Additional regulatory disclosures may be found here: > https://www.centrica.com/privacy-cookies-and-legal-disclaimer#email > > PH Jones is a trading name of British Gas Social Housing Limited. British > Gas Social Housing Limited (company no: 01026007), British Gas Trading > Limited (company no: 03078711), British Gas Services Limited (company no: > 3141243), British Gas Insurance Limited (company no: 06608316), British Gas > New Heating Limited (company no: 06723244), British Gas Services > (Commercial) Limited (company no: 07385984) and Centrica Energy (Trading) > Limited (company no: 02877397) are all wholly owned subsidiaries of > Centrica plc (company no: 3033654). Each company is registered in England > and Wales with a registered office at Millstream, Maidenhead Road, Windsor, > Berkshire SL4 5GD. > > British Gas Insurance Limited is authorised by the Prudential Regulation > Authority and regulated by the Financial Conduct Authority and the > Prudential Regulation Authority. British Gas Services Limited and Centrica > Energy (Trading) Limited are authorised and regulated by the Financial > Conduct Authority. British Gas Trading Limited is an appointed > representative of British Gas Services Limited which is authorised and > regulated by the Financial Conduct Authority. > > > > > The information contained in or attached to this email is intended only > for the use of the individual or entity to which it is addressed. If you > are not the intended recipient, or a person responsible for delivering it > to the intended recipient, you are not authorised to and must not disclose, > copy, distribute, or retain this message or any part of it. It may contain > information which is confidential and/or covered by legal professional or > other privilege under applicable law. > > The views expressed in this email are not necessarily the views of > Centrica plc or its subsidiaries, and the company, its directors, officers > or employees make no representation or accept any liability for its > accuracy or completeness unless expressly stated to the contrary. > > Additional regulatory disclosures may be found here: > https://www.centrica.com/privacy-cookies-and-legal-disclaimer#email > > PH Jones is a trading name of British Gas Social Housing Limited. British > Gas Social Housing Limited (company no: 01026007), British Gas Trading > Limited (company no: 03078711), British Gas Services Limited (company no: > 3141243), British Gas Insurance Limited (company no: 06608316), British Gas > New Heating Limited (company no: 06723244), British Gas Services > (Commercial) Limited (company no: 07385984) and Centrica Energy (Trading) > Limited (company no: 02877397) are all wholly owned subsidiaries of > Centrica plc (company no: 3033654). Each company is registered in England > and Wales with a registered office at Millstream, Maidenhead Road, Windsor, > Berkshire SL4 5GD. > > British Gas Insurance Limited is authorised by the Prudential Regulation > Authority and regulated by the Financial Conduct Authority and the > Prudential Regulation Authority. British Gas Services Limited and Centrica > Energy (Trading) Limited are authorised and regulated by the Financial > Conduct Authority. British Gas Trading Limited is an appointed > representative of British Gas Services Limited which is authorised and > regulated by the Financial Conduct Authority. > > > The information contained in or attached to this email is intended only > for the use of the individual or entity to which it is addressed. If you > are not the intended recipient, or a person responsible for delivering it > to the intended recipient, you are not authorised to and must not disclose, > copy, distribute, or retain this message or any part of it. It may contain > information which is confidential and/or covered by legal professional or > other privilege under applicable law. > > The views expressed in this email are not necessarily the views of > Centrica plc or its subsidiaries, and the company, its directors, officers > or employees make no representation or accept any liability for its > accuracy or completeness unless expressly stated to the contrary. > > Additional regulatory disclosures may be found here: > https://www.centrica.com/privacy-cookies-and-legal-disclaimer#email > > PH Jones is a trading name of British Gas Social Housing Limited. British > Gas Social Housing Limited (company no: 01026007), British Gas Trading > Limited (company no: 03078711), British Gas Services Limited (company no: > 3141243), British Gas Insurance Limited (company no: 06608316), British Gas > New Heating Limited (company no: 06723244), British Gas Services > (Commercial) Limited (company no: 07385984) and Centrica Energy (Trading) > Limited (company no: 02877397) are all wholly owned subsidiaries of > Centrica plc (company no: 3033654). Each company is registered in England > and Wales with a registered office at Millstream, Maidenhead Road, Windsor, > Berkshire SL4 5GD. > > British Gas Insurance Limited is authorised by the Prudential Regulation > Authority and regulated by the Financial Conduct Authority and the > Prudential Regulation Authority. British Gas Services Limited and Centrica > Energy (Trading) Limited are authorised and regulated by the Financial > Conduct Authority. British Gas Trading Limited is an appointed > representative of British Gas Services Limited which is authorised and > regulated by the Financial Conduct Authority. >
