Good intuition... yarn.scheduler.minimum-
allocation-mbis set to 2048 :)
I specified -ytm 2048 and -yjm 2048 and the job started right away, I will also try again later to see if it's not luck. Thanks a lot !
Regarding the version, it is still 0.1, and that I have no clue.... I downloaded 1.5.2 from this link : https://www.apache.org/dyn/
It should be the official build... It seems to be correct afterall, I see "1a9b648" there too: https://github.com/apache/
But I don't know why it's written version 0.1...
De : Gary Yao <gary@xxxxxxxxxxxxxxxxx>
Envoyé : mardi 7 août 2018 19:30
À : Florian Simond
Cc : user@xxxxxxxxxxxxxxxx
Objet : Re: Could not build the program from JAR file.Hi Florian,
Thank you for the logs. They look indeed strange but I cannot reproduce this
behavior. From the logs I can see that the ResourceManager is requesting
containers with different resource profiles (2048mb or 1024mb memory):
Requesting new TaskExecutor container with resources <memory:2048, vCores:1>. Number pending requests 1.
Requesting new TaskExecutor container with resources <memory:1024, vCores:1>. Number pending requests 1.
At the moment I am not exactly sure how this happens, and if this is
problematic at all. It would be helpful to know if you are configuring YARN
with a yarn.scheduler.minimum-
allocation-mb of 2048.
Here are some other things to try out for troubleshooting:
Can you try raising the the TM and JM memory to 2048mb (-ytm 2048 -yjm 2048)?
You are setting -ytm to 1024, which results in a heap size of only 424mb.
Is the deployment also slow if you are running example/streaming/WordCount.
The version in the log shows up as:
"Version: 0.1, Rev:1a9b648, Date:25.07.2018 @ 17:10:13 GMT".
This does not seem to be an official Flink 1.5.2 distribution. Is there are
reason for that, and can you rule out that there are no changes? Maybe try out
the official build?
On Tue, Aug 7, 2018 at 2:44 PM, Florian Simond <florian.simond@xxxxxxxxxx> wrote:
No, I am not starting multiple "per-job clusters".
I didn't configure anything regarding the number of slots per TM, so I guess the default value (1 then).
But on the YARN UI I see that the number of "running containers" varies a lot (13 then 1 then 8 then 2 then 27 then 6 etc...)
Here is the full jobmanager log:
This time it took longer to start (10 minutes)
And completed on this line:2018-08-07 14:31:11,852 INFO org.apache.flink.runtime.execu
tiongraph.ExecutionGraph - Source: Custom Source -> Sink: Unnamed (1/1) (655509c673d8ae19aac195276ad2c 3e6) switched from DEPLOYING to RUNNING.
Thanks a lot for your help and your time,
5 minutes sounds too slow. Are you starting multiple "per-job clusters" at the
same time? How many slots do you configure per TM? After you submit the job,
how many resources do you have left in your YARN cluster?
It might be that you are affected by FLINK-9455 : Flink requests
unnecessary resources from YARN and blocks the execution of other jobs
temporarily. The workaround is to configure only one slot per TM.
If the above does not help, can you attach the full ClusterEntrypoint
On Tue, Aug 7, 2018 at 12:34 PM, Florian Simond <florian.simond@xxxxxxxxxx> wrote:
So it is now normal that it takes around 5 minutes to start processing ? The job is reading from kafka and writing back into another kafka topic. When I start the job, it takes roughly 5 minutes before I get something in the output topic.
I see a lot of2018-08-07 12:20:34,672 INFO org.apache.flink.yarn.YarnReso
urceManager - Received new container: XXX - Remaining pending container requests: 0 2018-08-07 12:20:34,672 INFO org.apache.flink.yarn.YarnReso urceManager - Returning excess container XXX.I see a lot of those lines during the first five minutes.
I'm not sure I need to have a static set of TMs, but as we have a limited set of nodes, and several jobs, it could be harder to make sure they do not interfere with each other...
You can find more information about the re-worked deployment model here:
TaskManagers are started and shut down according to the slot requirements of
the jobs. It is possible to return to the old behavior by setting
in flink-conf.yaml. However, this mode is deprecated and will be removed soon.
Can you explain why you need to have a static set of TMs?
On Tue, Aug 7, 2018 at 12:07 PM, Florian Simond <florian.simond@xxxxxxxxxx> wrote:
Indeed, that's the solution.
It was done automatically before with 1.4.2, that's why I missed that part...
Do you have any pointer about the dynamic number of TaskManagers ? I'm curious to know how it works. Is it possible to still fix it ?
Can you run
export HADOOP_CLASSPATH=`hadoop classpath`
before submitting the job ?
Moreover, you should not use the -yn parameter. Beginning with Flink 1.5, the
number of TaskManagers is not fixed anymore.
/flink/flink-docs-release-1.5/ ops/deployment/hadoop.html#con figuring-flink-with-hadoop-cla sspaths
On Tue, Aug 7, 2018 at 9:22 AM, Florian Simond <florian.simond@xxxxxxxxxx> wrote:
In the log, I can see that:
First exception is a warning, not sure if it is important.
Second one seems to be the one. It tries to find the file "-yn" ???
2018-08-07 09:16:04,776 WARN org.apache.flink.client.cli.Cl
iFrontend - Could not load CLI class org.apache.flink.yarn.cli.Flin kYarnSessionCli.java.lang.NoClassDefFoundError : org/apache/hadoop/conf/Configu rationat java.lang.Class.forName0(Nativ e Method)at java.lang.Class.forName(Class. java:264)at org.apache.flink.client.cli.Cl iFrontend.loadCustomCommandLin e(CliFrontend.java:1208)at org.apache.flink.client.cli.Cl iFrontend.loadCustomCommandLin es(CliFrontend.java:1164)at org.apache.flink.client.cli.Cl iFrontend.main(CliFrontend.jav a:1090)Caused by: java.lang.ClassNotFoundExcepti on: org.apache.hadoop.conf.Configu rationat java.net.URLClassLoader.findCl ass(URLClassLoader.java:381)at java.lang.ClassLoader.loadClas s(ClassLoader.java:424)at sun.misc.Launcher$AppClassLoad er.loadClass(Launcher.java:335 )at java.lang.ClassLoader.loadClas s(ClassLoader.java:357)... 5 more2018-08-07 09:16:04,789 INFO org.apache.flink.core.fs.FileS ystem - Hadoop is not in the classpath/dependencies. The extended set of supported File Systems via Hadoop is not available.2018-08-07 09:16:04,967 INFO org.apache.flink.runtime.secur ity.modules.HadoopModuleFactor y - Cannot create Hadoop Security Module because Hadoop cannot be found in the Classpath.2018-08-07 09:16:04,991 INFO org.apache.flink.runtime.secur ity.SecurityUtils - Cannot install HadoopSecurityContext because Hadoop cannot be found in the Classpath.2018-08-07 09:16:05,041 INFO org.apache.flink.client.cli.Cl iFrontend - Running 'run' command.2018-08-07 09:16:05,046 INFO org.apache.flink.client.cli.Cl iFrontend - Building program from JAR file2018-08-07 09:16:05,046 ERROR org.apache.flink.client.cli.Cl iFrontend - Invalid command line arguments.org.apache.flink.client.cli.Cl iArgsException: Could not build the program from JAR file.at org.apache.flink.client.cli.Cl iFrontend.run(CliFrontend.java :208)at org.apache.flink.client.cli.Cl iFrontend.parseParameters(CliF rontend.java:1025)at org.apache.flink.client.cli.Cl iFrontend.lambda$main$9(CliFro ntend.java:1101)at org.apache.flink.runtime.secur ity.NoOpSecurityContext.runSec ured(NoOpSecurityContext.java: 30)at org.apache.flink.client.cli.Cl iFrontend.main(CliFrontend.jav a:1101)Caused by: java.io.FileNotFoundException: JAR file does not exist: -ynat org.apache.flink.client.cli.Cl iFrontend.buildProgram(CliFron tend.java:828)at org.apache.flink.client.cli.Cl iFrontend.run(CliFrontend.java :205)... 4 more
The error message is because of a FileNotFoundException, see here. Is there any more information about the exception. Do you make sure the jar exist?
/flink/blob/master/flink-clien ts/src/main/java/org/apache/fl ink/client/cli/CliFrontend.jav a#L209
2018-08-07 14:28 GMT+08:00 Gary Yao <gary@xxxxxxxxxxxxxxxxx>:
You write that Flink 1.4.2 works but what version is not working for you?
On Tue, Aug 7, 2018 at 8:25 AM, Florian Simond <florian.simond@xxxxxxxxxx> wrote:
I'm trying to run the wordCount example on my YARN cluster and this is not working.. I get the error message specified in title: Could not build the program from JAR file.
> $ ./bin/flink run -m yarn-cluster -yn 4 -yjm 1024 -ytm 4096 ./examples/batch/WordCount.jar> Setting HADOOP_CONF_DIR=/etc/hadoop/co
nf because no HADOOP_CONF_DIR was set.> Could not build the program from JAR file.
> Use the help option (-h or --help) to get help on the command.
I also have the same problem with a custom JAR...
With Flink 1.4.2, I have no problem at all. Both the WordCount example and my custom JAR are working...
What do I do wrong ?