That's the culprit and Alexey's suggested shading resolved the problem
for me.
Max and Alexey, thanks for all you help on this one!
Wayne
On 2016-12-05 11:02 AM, Alexey Demin wrote:
Hi Max,
Optional:
file:/opt/cloudera/parcels/CDH-5.9.0-1.cdh5.9.0.p0.23/jars/guava-11.0.2.jar
Absent:
file:/opt/cloudera/parcels/CDH-5.9.0-1.cdh5.9.0.p0.23/jars/guava-11.0.2.jar
Step for reproducing:
1) CDH hadoop cluster
2) download and unpack flink-1.1.3-bin-hadoop26-scala_2.10.tgz
3) create YARN container: HADOOP_USER_NAME=flink
YARN_CONF_DIR=yarn-conf/ ./bin/yarn-session.sh -n 1 -jm 1024 -tm 1024
-z /pr/test1
4) set ApplicationMaster and make one shaded jar
5) execute: java -cp target/example.jar ClassloaderTest
p.s. source code for validation:
https://gist.github.com/xhumanoid/83e45ff8a9209b1d7115bda33d18c966
Thanks,
Alexey Diomin
2016-12-05 18:55 GMT+04:00 Maximilian Michels <[email protected]
<mailto:[email protected]>>:
Hi Alexey,
I suspected that Hadoop's classpath which is added after the main
Flink classpath might sneak in an older version of Guava. However,
when your jar is executed, the Flink class loader will make sure that
your classes (including Beam's) will be loaded first. Normally,
Hadoop's Guava would not make it in that way because it comes later in
the classpath. However, there's probably some Hadoop code executed
before, that loads the Guava classes.
It is problematic that the Kafka module of Beam does not shade Guava
when you use it in combination with Hadoop which does not shade Guava
either. I couldn't manage to reproduce it on my cluster setup. The fix
which Alexey suggested, should make it work because it relocates your
Guava classes and thereby avoids a clash with Hadoop's version.
I would still be interested in the output of the code I posted earlier
on your cluster. Just to confirm the suspicion.
Thanks,
Max
On Mon, Dec 5, 2016 at 2:22 PM, Alexey Demin <[email protected]
<mailto:[email protected]>> wrote:
> Hi Max
>
> No, it's main problem in flink runner
> I investigate problem and found than AppClassLoader load this
classes:
>
> com.google.common.base.Absent
> com.google.common.base.Function
> com.google.common.base.Optional
> com.google.common.base.Optional$Absent
> com.google.common.base.Optional$Present
>
> But Optional$Absent existed only in old version guava, after
commit "Make
> Optional GWT serializable"
3e82ff72b19d4230eba795b2760036eb18dfd4ff guava
> split Optional on 3 different classes and made constructor for
Optional
> package private. Optional$Absent and Absent can be loaded
together from
> different versions of guava, but can't work.
>
> but because FlinkUserCodeClassLoader have parent on
AppClassLoader we have
> problem with guava versions from hadoop classpath in cloudera
distributions.
>
> My config:
> FlinkPipelineOptions options =
> PipelineOptionsFactory.as(FlinkPipelineOptions.class);
> options.setRunner(FlinkRunner.class);
> options.setFlinkMaster("<host:port>");
>
> and execution command: java -jar runner-shaded.jar
>
> Flink running on cloudera yarn.
>
> My small hotfix it's addition configuration shaded plugin for my
> application:
>
> <relocations>
> <relocation>
> <pattern>com.google.common</pattern>
> <shadedPattern>my.shaded.com
<http://my.shaded.com>.google.common</shadedPattern>
> </relocation>
> </relocations>
>
> It remove conflict with AppClassLoader version of guava and
alway use
> correct version from my application.
>
> Thanks,
> Alexey Diomin
>
>
> 2016-12-05 14:50 GMT+04:00 Maximilian Michels <[email protected]
<mailto:[email protected]>>:
>>
>> Hi Wayne,
>>
>> That seems like you ran that from the IDE. I'm interested how
the output
>> would look like on the cluster where you experienced your problems.
>>
>> Thanks,
>> Max
>>
>> On Fri, Dec 2, 2016 at 7:27 PM, Wayne Collins <[email protected]
<mailto:[email protected]>> wrote:
>>>
>>> Hi Max,
>>>
>>> Here's the output:
>>> ---------------------
>>> Optional:
>>>
file:/home/wayneco/workspace/beam-starter/beam-starter/./target/beam-starter-0.2.jar
>>> Absent:
>>>
file:/home/wayneco/workspace/beam-starter/beam-starter/./target/beam-starter-0.2.jar
>>> ---------------------
>>>
>>> Thanks for your help!
>>> Wayne
>>>
>>>
>>>
>>> On 2016-12-02 08:42 AM, Maximilian Michels wrote:
>>>>
>>>> Hi Wayne,
>>>>
>>>> Thanks for getting back to me. Could you compile a new
version of your
>>>> job with the following in your main method?
>>>>
>>>> URL location1 =
>>>>
>>>>
com.google.common.base.Optional.class.getProtectionDomain().getCodeSource().getLocation();
>>>> System.out.println("Optional: " + location1);
>>>> URL location2 =
>>>>
>>>>
Class.forName("com.google.common.base.Optional").getProtectionDomain().getCodeSource().getLocation();
>>>> System.out.println("Absent: " + location2);
>>>>
>>>>
>>>> Could you run this on your cluster node with the flink
command? This
>>>> should give us a hint from where the Guava library is
bootstrapped.
>>>>
>>>> Thanks,
>>>> Max
>>>>
>>>>
>>>> On Thu, Dec 1, 2016 at 7:54 PM, Wayne Collins
<[email protected] <mailto:[email protected]>> wrote:
>>>>>
>>>>> Hi Max,
>>>>>
>>>>> Here is the result from the "flink run" launcher node (devbox):
>>>>> -----------------------
>>>>> root@devbox:~# echo
>>>>>
>>>>>
${HADOOP_CLASSPATH}:${HADOOP_CONF_DIR}:${YARN_CONF_DIR}:${HBASE_CONF_DIR}
>>>>> :/etc/hadoop-conf:/etc/yarn-conf:
>>>>> -----------------------
>>>>>
>>>>> Here is the result from one of the Cloudera YARN nodes as root:
>>>>> -----------------------
>>>>> [root@hadoop0 ~]# echo
>>>>>
>>>>>
${HADOOP_CLASSPATH}:${HADOOP_CONF_DIR}:${YARN_CONF_DIR}:${HBASE_CONF_DIR}
>>>>> :::
>>>>> -----------------------
>>>>>
>>>>> Here is the result from one of the Cloudera YARN nodes as yarn:
>>>>> -----------------------
>>>>> [yarn@hadoop0 ~]$ echo
>>>>>
>>>>>
${HADOOP_CLASSPATH}:${HADOOP_CONF_DIR}:${YARN_CONF_DIR}:${HBASE_CONF_DIR}
>>>>> :::
>>>>> -----------------------
>>>>>
>>>>>
>>>>> Note that both the yarn-session.sh and the flink run
commands are run
>>>>> as
>>>>> root on devbox.
>>>>>
>>>>> Software version details:
>>>>> devbox has these versions of the client software:
>>>>> flink-1.1.2
>>>>> hadoop-2.6.0
>>>>> kafka_2.11-0.9.0.1
>>>>> (also reproduced the problem with kafka_2.10-0.9.0.1)
>>>>>
>>>>> The cluster (providing YARN) is:
>>>>> CDH5 - 5.8.2-1.cdh5.8.2.p0.3 (Hadoop 2.6.0)
>>>>> Kafka - 2.0.2-1.2.0.2.p0.5 (Kafka 0.9.0)
>>>>>
>>>>> Thanks for your help!
>>>>> Wayne
>>>>>
>>>>>
>>>>>
>>>>> On 2016-12-01 12:54 PM, Maximilian Michels wrote:
>>>>>
>>>>> What is the output of the following on the nodes? I have a
suspision
>>>>> that something sneaks in from one of the classpath variables
that
>>>>> Flink picks up:
>>>>>
>>>>> echo
>>>>>
>>>>>
${HADOOP_CLASSPATH}:${HADOOP_CONF_DIR}:${YARN_CONF_DIR}:${HBASE_CONF_DIR}
>>>>>
>>>>> On Tue, Nov 29, 2016 at 9:17 PM, Wayne Collins
<[email protected] <mailto:[email protected]>>
>>>>> wrote:
>>>>>
>>>>> Hi Max,
>>>>>
>>>>> I rebuilt my sandbox with Beam 0.3.0-incubating and Flink
1.1.2 and I'm
>>>>> still seeing the following error message with the
StreamWordCount demo
>>>>> code:
>>>>>
>>>>> Caused by: java.lang.IllegalAccessError: tried to access method
>>>>> com.google.common.base.Optional.<init>()V from class
>>>>> com.google.common.base.Absent
>>>>> at com.google.common.base.Absent.<init>(Absent.java:35)
>>>>> at
com.google.common.base.Absent.<clinit>(Absent.java:33)
>>>>> at sun.misc.Unsafe.ensureClassInitialized(Native
Method)
>>>>> ...
>>>>>
>>>>>
>>>>> (snip)
>>>
>>>
>>
>
