java.lang.Exception: TaskManager was lost/killed

[Lasse Nedergaard]

Did a test where all our job used FSStateBackend with the same result task managers where killed with memory exception.

Based on that I will conclude that either does both RocksDB and FsStateBackend leaks, or Flink does when jobs restarts

Lasse

2018-04-12 13:56 GMT+02:00 Lasse Nedergaard <[hidden email]>:

Hi.

I have news, not sure it is good.

I have tried to limit RocksDb with the configuration mention from the links provided, but it didn't solved the problem. (se the configuration below.).

I build a new cluster with 19 GB mem allocated, 8 vcores and thereby 8 slots. Out heap is around 6 GB and non-heap around 1.6 GB, so plenty of free memory but our task managers still get killed with 

2018-04-12 11:18:15,807 INFO org.apache.flink.yarn.YarnFlinkResourceManager - Diagnostics for container container_1523368093942_0004_01_000073 in state COMPLETE : exitStatus=Pmem limit exceeded (-104) diagnostics=Container [pid=13279,containerID=container_1523368093942_0004_01_000073] is running beyond physical memory limits. Current usage: 19.6 GB of 19.5 GB physical memory used; 30.7 GB of 97.7 GB virtual memory used. Killing container.

I also did a heap dump on the container and it shows the same utilisation as Flink metrics regarding memory use.

This morning we played around with our Cassandra cluster and we lost qurum and then our jobs inserting into Cassandra failed all the time and get restarted because of this. Right away the frequencies of task manager kills increased. 

Looking at https://issues.apache.org/jira/browse/FLINK-7289 explains the same problem that a restart cause a leak (not sure where, and who should clean it up) and therefore we see our task managers get killed.

It also make sense  as our new pipeline fails more than the old one, as we get exceptions from AWS kinesis producer, and the old pipeline didn't use kinesis producer. Problem described here https://github.com/awslabs/amazon-kinesis-producer/issues/39#issuecomment-379759277

So my conclusion right now is that Flink on EMR doesn't manage to restarts job without leaks, and we therefore can't rely on restart and have to protect us as much as possible against external errors. Not always an easy task. 

Next I will try is to exchange RocksDb with FsStateBackend to see if it's RockDbs that cause the problem, or something else.

Please let me know it have have miss anything.

Lasse Nedergaard

Rocks db configuration

final long blockCacheSize = 64 * SizeUnit.MB;
final long blockSize = 4 * SizeUnit.KB;
final long max_buffer_size = 128 * SizeUnit.MB;
final int max_write_buffer_number = 5; //default 2
final int minimum_buffer_number_to_merge = 2; //default 2
//If you set cache_index_and_filter_blocks to true, index and filter blocks will be stored in block cache, together with all other data blocks
final boolean cacheIndexAndFilterBlocks = true;
//With this option turned on, we will not build bloom filters on the last level, which contains 90% of the database. Thus, the memory usage for bloom filters will be 10X less. You will pay one IO for each Get() that doesn't find data in the database
final boolean optimizeFilterForHits = true;

PredefinedOptions baseOptions = PredefinedOptions.FLASH_SSD_OPTIMIZED;

baseOptions.createDBOptions()
        .setMaxBackgroundCompactions(4);
baseOptions.createColumnOptions()
        .setWriteBufferSize(max_buffer_size)
        .setMaxBytesForLevelBase(max_buffer_size)
        .setMaxWriteBufferNumber(max_write_buffer_number)
        .setMinWriteBufferNumberToMerge(minimum_buffer_number_to_merge)
        .setOptimizeFiltersForHits(optimizeFilterForHits)
        .setTableFormatConfig(new BlockBasedTableConfig()
                .setBlockCacheSize(blockCacheSize)
                .setBlockSize(blockSize)
                .setCacheNumShardBits(6)
                .setBlockSizeDeviation(5)
                .setBlockRestartInterval(10)
                .setCacheIndexAndFilterBlocks(cacheIndexAndFilterBlocks)
                .setHashIndexAllowCollision(false)
                .setBlockCacheCompressedSize(64 * SizeUnit.KB)
                .setBlockCacheCompressedNumShardBits(10))
;

2018-04-11 16:27 GMT+02:00 Ken Krugler <[hidden email]>:

I’ve been able to hook up YourKit (commercial Java profiler) to tasks running in EMR.

IIRC there was a bit of pain to get the right port open (AWS security group configuration fun)

— Ken

On Apr 11, 2018, at 2:38 AM, Lasse Nedergaard <[hidden email]> wrote:

Hi.

Thanks for the information I have taked a look at it and tried to change some of the RocksDBConfiguration, and restarted all our jobs. I let you all know when I have figured it all out.

Anyone that from top of your head can tell me how I can see the memory usage inside a Yarn container running on EMR? 

I need to go into these details as I yesterday created a new cluster with task-managers allocated 20 GB mem, and they was killed during the night and the heap and non-heap was way below the allocated memory (see attatchment) so I would like to figure out what is using all the memory, RocksDB, aws Kinesis procedure daemon or something else.

Lasse Nedergaard 

2018-04-11 04:38:28,105 INFO org.apache.flink.yarn.YarnFlinkResourceManager - Diagnostics for container container_1523368093942_0004_01_000006 in state COMPLETE : exitStatus=Pmem limit exceeded (-104) diagnostics=Container [pid=14155,containerID=container_1523368093942_0004_01_000006] is running beyond physical memory limits. Current usage: 19.5 GB of 19.5 GB physical memory used; 29.5 GB of 97.7 GB virtual memory used. Killing container. Dump of the process-tree for container_1523368093942_0004_01_000006 : |- PID PPID PGRPID SESSID CMD_NAME USER_MODE_TIME(MILLIS) SYSTEM_TIME(MILLIS) VMEM_USAGE(BYTES) RSSMEM_USAGE(PAGES) FULL_CMD_LINE |- 14173 14155 14155 14155 (java) 731856 128928 31159238656 5120164 /usr/lib/jvm/java-openjdk/bin/java -Xms15000m -Xmx15000m -Dlog.file=/var/log/hadoop-yarn/containers/application_1523368093942_0004/container_1523368093942_0004_01_000006/taskmanager.log -Dlogback.configurationFile=file:./logback.xml -Dlog4j.configuration=file:./log4j.properties org.apache.flink.yarn.YarnTaskManager --configDir . |- 14155 14153 14155 14155 (bash) 0 0 115818496 664 /bin/bash -c /usr/lib/jvm/java-openjdk/bin/java -Xms15000m -Xmx15000m -Dlog.file=/var/log/hadoop-yarn/containers/application_1523368093942_0004/container_1523368093942_0004_01_000006/taskmanager.log -Dlogback.configurationFile=file:./logback.xml -Dlog4j.configuration=file:./log4j.properties org.apache.flink.yarn.YarnTaskManager --configDir . 1> /var/log/hadoop-yarn/containers/application_1523368093942_0004/container_1523368093942_0004_01_000006/taskmanager.out 2> /var/log/hadoop-yarn/containers/application_1523368093942_0004/container_1523368093942_0004_01_000006/taskmanager.err |- 3579 14173 14155 14155 (kinesis_produce) 150 25 406056960 3207 /tmp/amazon-kinesis-producer-native-binaries/kinesis_producer_d93825f806782576ef9f09eef67a2baeadfec35c -o /tmp/amazon-kinesis-producer-native-binaries/amz-aws-kpl-out-pipe-f8b715df -i /tmp/amazon-kinesis-producer-native-binaries/amz-aws-kpl-in-pipe-f787eee9 -c 0800328401080110FFFFFFFF0F18809003220028BB0330F403388080C00240F02E480050005A0060BB036A04696E666F70187A05736861726482010864657461696C65648A01164B696E6573697350726F64756365724C6962726172799001E0D403980101A00164A80164B001B0EA01BA010965752D776573742D31C001F02EC80101D00101D8010A -k 08FFFFFFFFFFFFFFFF7F4A44080012400A14414B49414A4F52573753514F355143324B57414112286F78693674774B524A3243616665464C4F7535304F2B506B597369562B47545368526C762B625932 -t -w 08FFFFFFFFFFFFFFFF7F4A44080112400A14414B49414A4F52573753514F355143324B57414112286F78693674774B524A3243616665464C4F7535304F2B506B597369562B47545368526C762B625932 Container killed on request. Exit code is 143 Container exited with a non-zero exit code 143

2018-04-11 5:03 GMT+02:00 周思华 <[hidden email]>:

Hi Lasse,

I met that before. I think maybe the non-heap memory trend of the graph you attached is the "expected" result ... Because rocksdb will keep the a "filter (bloom filter)" in memory for every opened sst file by default, and the num of the sst file will increase by time, so it looks like a leak. There is a issue(https://issues.apache.org/jira/browse/FLINK-7289) Stefan created to track this, and the page(https://github.com/facebook/rocksdb/wiki/Memory-usage-in-RocksDB) from RocksDB's wiki could give you a better understand of the memory used by RocksDB, and Stefan please correct me if I bring any wrong information above.

Best Regards,

Sihua Zhou

On 04/11/2018 09:55，[hidden email] wrote：

Please see the last comment on this issue:

https://github.com/facebook/rocksdb/issues/3216

FYI

On Tue, Apr 10, 2018 at 12:25 AM, Lasse Nedergaard <[hidden email]> wrote:

<Screen Shot 2018-04-10 at 09.09.47.png>
This graph shows Non-Heap . If the same pattern exists it make sense that it will try to allocate more memory and then exceed the limit. I can see the trend for all other containers that has been killed. So my question is now, what is using non-heap memory?

From http://mail-archives.apache.org/mod_mbox/flink-user/201707.mbox/%3CCANC1h_u0dQQvbysDAoLLbEmeWaxiimTMFjJCribPfpo0iDLZ_g@...%3E it look like RockDb could be guilty.

I have job using incremental checkpointing and some without, some optimised for FLASH_SSD. all have same pattern

Lasse 

2018-04-10 8:52 GMT+02:00 Lasse Nedergaard <[hidden email]>:

Hi.

I found the exception attached below, for our simple job. It states that our task-manager was killed du to exceed memory limit on 2.7GB. But when I look at Flink metricts just 30 sec before it use 1.3 GB heap and 712 MB Non-Heap around 2 GB. 

So something else are also using memory inside the conatianer any idea how to figure out what?

As a side note we use RockDBStateBackend with this configuration

env.getCheckpointConfig().setMinPauseBetweenCheckpoints((long)(config.checkPointInterval * 0.75));
env.enableCheckpointing(config.checkPointInterval, CheckpointingMode.AT_LEAST_ONCE);
env.setStateBackend(new RocksDBStateBackend(config.checkpointDataUri));

Where checkpointDataUri point to S3

Lasse Nedergaard

2018-04-09 16:52:01,239 INFO  org.apache.flink.yarn.YarnFlinkResourceManager                - Diagnostics for container container_1522921976871_0001_01_000079 in state COMPLETE : exitStatus=Pmem limit exceeded (-104) diagnostics=Container [pid=30118,containerID=container_1522921976871_0001_01_000079] is running beyond physical memory limits. Current usage: 2.7 GB of 2.7 GB physical memory used; 4.9 GB of 13.4 GB virtual memory used. Killing container.

Dump of the process-tree for container_1522921976871_0001_01_000079 :

|- PID PPID PGRPID SESSID CMD_NAME USER_MODE_TIME(MILLIS) SYSTEM_TIME(MILLIS) VMEM_USAGE(BYTES) RSSMEM_USAGE(PAGES) FULL_CMD_LINE

|- 30136 30118 30118 30118 (java) 245173 68463 5193723904 703845 /usr/lib/jvm/java-openjdk/bin/java -Xms2063m -Xmx2063m -Dlog.file=/var/log/hadoop-yarn/containers/application_1522921976871_0001/container_1522921976871_0001_01_000079/taskmanager.log -Dlogback.configurationFile=file:./logback.xml -Dlog4j.configuration=file:./log4j.properties org.apache.flink.yarn.YarnTaskManager --configDir . 

|- 30118 30116 30118 30118 (bash) 0 0 115818496 674 /bin/bash -c /usr/lib/jvm/java-openjdk/bin/java -Xms2063m -Xmx2063m  -Dlog.file=/var/log/hadoop-yarn/containers/application_1522921976871_0001/container_1522921976871_0001_01_000079/taskmanager.log -Dlogback.configurationFile=file:./logback.xml -Dlog4j.configuration=file:./log4j.properties org.apache.flink.yarn.YarnTaskManager --configDir . 1> /var/log/hadoop-yarn/containers/application_1522921976871_0001/container_1522921976871_0001_01_000079/taskmanager.out 2> /var/log/hadoop-yarn/containers/application_1522921976871_0001/container_1522921976871_0001_01_000079/taskmanager.err 

2018-04-09 16:51:26,659 DEBUG org.trackunit.tm2.LogReporter                                 - gauge.ip-10-1-1-181.taskmanager.container_1522921976871_0001_01_000079.Status.JVM.Memory.Heap.Used=1398739496

2018-04-09 16:51:26,659 DEBUG org.trackunit.tm2.LogReporter                                 - gauge.ip-10-1-1-181.taskmanager.container_1522921976871_0001_01_000079.Status.JVM.Memory.NonHeap.Used=746869520

 

2018-04-09 23:52 GMT+02:00 Ken Krugler <[hidden email]>:

Hi Chesnay,

Don’t know if this helps, but I’d run into this as well, though I haven’t hooked up YourKit to analyze exactly what’s causing the memory problem.

E.g. after about 3.5 hours running locally, it failed with memory issues.

In the TaskManager logs, I start seeing exceptions in my code….

java.lang.OutOfMemoryError: GC overhead limit exceeded

And then eventually...

2018-04-07 21:55:25,686 WARN  org.apache.flink.runtime.accumulators.AccumulatorRegistry     - Failed to serialize accumulators for task.
java.lang.OutOfMemoryError: GC overhead limit exceeded

Immediately after this, one of my custom functions gets a close() call, and I see a log msg about it "switched from RUNNING to FAILED”.

After this, I see messages that the job is being restarted, but the TaskManager log output abruptly ends.

In the Job Manager log, this is what is output following the time of the last TaskManager logging output:

2018-04-07 21:57:33,702 INFO  org.apache.flink.runtime.checkpoint.CheckpointCoordinator     - Triggering checkpoint 129 @ 1523163453702
2018-04-07 21:58:43,916 WARN  akka.remote.ReliableDeliverySupervisor                        - Association with remote system [akka.tcp://[hidden email]1.ca.comcast.net:63780] has failed, address is now gated for [5000] ms. Reason: [Disassociated] 
2018-04-07 21:58:51,084 WARN  akka.remote.transport.netty.NettyTransport                    - Remote connection to [null] failed with java.net.ConnectException: Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780
2018-04-07 21:58:51,086 WARN  akka.remote.ReliableDeliverySupervisor                        - Association with remote system [akka.tcp://flink@...1.ca.comcast.net:63780] has failed, address is now gated for [5000] ms. Reason: [Association failed with [akka.tcp://flink@...1.ca.comcast.net:63780]] Caused by: [Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780]
2018-04-07 21:59:01,047 WARN  akka.remote.transport.netty.NettyTransport                    - Remote connection to [null] failed with java.net.ConnectException: Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780
2018-04-07 21:59:01,050 WARN  akka.remote.ReliableDeliverySupervisor                        - Association with remote system [akka.tcp://flink@...1.ca.comcast.net:63780] has failed, address is now gated for [5000] ms. Reason: [Association failed with [akka.tcp://flink@...1.ca.comcast.net:63780]] Caused by: [Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780]
2018-04-07 21:59:11,057 WARN  akka.remote.ReliableDeliverySupervisor                        - Association with remote system [akka.tcp://flink@...1.ca.comcast.net:63780] has failed, address is now gated for [5000] ms. Reason: [Association failed with [akka.tcp://flink@...1.ca.comcast.net:63780]] Caused by: [Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780]
2018-04-07 21:59:11,058 WARN  akka.remote.transport.netty.NettyTransport                    - Remote connection to [null] failed with java.net.ConnectException: Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780
2018-04-07 21:59:21,049 WARN  akka.remote.transport.netty.NettyTransport                    - Remote connection to [null] failed with java.net.ConnectException: Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780
2018-04-07 21:59:21,049 WARN  akka.remote.ReliableDeliverySupervisor                        - Association with remote system [akka.tcp://flink@...1.ca.comcast.net:63780] has failed, address is now gated for [5000] ms. Reason: [Association failed with [akka.tcp://flink@...1.ca.comcast.net:63780]] Caused by: [Connection refused: kens-mbp.hsd1.ca.comcast.net/192.168.3.177:63780]
2018-04-07 21:59:21,056 WARN  akka.remote.RemoteWatcher                                     - Detected unreachable: [akka.tcp://flink@...1.ca.comcast.net:63780]
2018-04-07 21:59:21,063 INFO  org.apache.flink.runtime.jobmanager.JobManager                - Task manager akka.tcp://[e-mail]:63780/user/taskmanager terminated.
2018-04-07 21:59:21,064 INFO  org.apache.flink.runtime.executiongraph.ExecutionGraph        - FetchUrlsFunction for sitemap -> ParseSiteMapFunction -> OutlinkToStateUrlFunction (1/1) (3e9374d1bf5fdb359e3a624a4d5d659b) switched from RUNNING to FAILED.
java.lang.Exception: TaskManager was lost/killed: c51d3879b6244828eb9fc78c943007ad @ kens-mbp.hsd1.ca.comcast.net (dataPort=63782)

— Ken

On Apr 9, 2018, at 12:48 PM, Chesnay Schepler <[hidden email]> wrote:

We will need more information to offer any solution. The exception simply means that a TaskManager shut down, for which there are a myriad of possible explanations.

Please have a look at the TaskManager logs, they may contain a hint as to why it shut down.

On 09.04.2018 16:01, Javier Lopez wrote:

Hi,

"are you moving the job  jar to  the ~/flink-1.4.2/lib path ?  " -> Yes, to every node in the cluster.

On 9 April 2018 at 15:37, miki haiat <[hidden email]> wrote:

Javier 

"adding the jar file to the /lib path of every task manager"
are you moving the job  jar to  the ~/flink-1.4.2/lib path ?  

On Mon, Apr 9, 2018 at 12:23 PM, Javier Lopez <[hidden email]> wrote:

Hi,

We had the same metaspace problem, it was solved by adding the jar file to the /lib path of every task manager, as explained here https://ci.apache.org/projects/flink/flink-docs-release-1.4/monitoring/debugging_classloading.html#avoiding-dynamic-classloading. As well we added these java options: "-XX:CompressedClassSpaceSize=100M -XX:MaxMetaspaceSize=300M -XX:MetaspaceSize=200M "

From time to time we have the same problem with TaskManagers disconnecting, but the logs are not useful. We are using 1.3.2.

On 9 April 2018 at 10:41, Alexander Smirnov <[hidden email]> wrote:

I've seen similar problem, but it was not a heap size, but Metaspace.

It was caused by a job restarting in a loop. Looks like for each restart, Flink loads new instance of classes and very soon in runs out of metaspace.

I've created a JIRA issue for this problem, but got no response from the development team on it: https://issues.apache.org/jira/browse/FLINK-9132

On Mon, Apr 9, 2018 at 11:36 AM 王凯 <[hidden email]> wrote:

thanks a lot,i will try it

在 2018-04-09 00:06:02，"TechnoMage" <[hidden email]> 写道：

I have seen this when my task manager ran out of RAM.  Increase the heap size.

flink-conf.yaml:

taskmanager.heap.mb 

jobmanager.heap.mb

Michael

On Apr 8, 2018, at 2:36 AM, 王凯 <[hidden email]> wrote:

<QQ图片20180408163927.png>

hi all, recently, i found a problem,it runs well when start. But after long run,the exception display as above,how can resolve it?

--------------------------------------------

http://about.me/kkrugler

[phone]

<Screen Shot 2018-04-11 at 11.36.32.png>

--------------------------------------------

http://about.me/kkrugler

[phone]