|Thanks for figuring this out, Shuyi!|
I think the bug is introduced in FLINK-7643 (Rework FileSystem loading to use factories). In YarnApplicationMasterRunner, after the JIRA, FileSystem was not properly initialized with the correct flink configuration before calling runApplicationMaster(). W/o the initialization, a call of FileSystem.get("hdfs://nameservice0/
") will cause the HadoopFsFactory to be initialized with an empty Flink Configuration, thus not able to recognize nameservice0.
I've composed unittests in YarnApplicationMasterRunner, YarnTaskManagerRunnerFactory and YarnTaskExecutorRunner, and verify that YarnApplicationMasterRunner fail to initialize the FileSystem properly, and YarnTaskManagerRunnerFactory does the right thing, so does YarnTaskExecutorRunner.
I'll create a JIRA to track the fix.
Thanks for looking into this! Yes, we toggled Flink 1.4 back to Flink 1.3.2 and it works. So seems truly a feature disparity between 1.3.2 and 1.4.
Another thing: when you retry this again with Flink 1.3.2 it works? I'm trying to rule out another problem in the setup.
This seems to be a bug in the FileSystems and how they're initialized. I'm looking into this myself but I'm also looping in Stephan and Stefan how have worked on this the most in the past. Maybe they have some valuable input.
We recently bumped to Flink 1.4 from 1.3.2, and found out an issue on HDFS configuration.
We are using FlinkYarnSessionCli to start the cluster and submit job.
In 1.3.2, we set below Flink properties when using checkpoints:
The mapping between logical nameservice (nameservice0) and actual namenodes hostports are passed to Flink via yarnship/core-site.xml (by providing the -yt option), and set fs.hdfs.hadoopconf=yarnship/
However, we encountered below error after bumping to 1.4, which caused the checkpointing to fail.
It does not recognize nameservice0 because the core-site.xml on the actual machine (read in by Flink via $HADOOP_CONF_DIR) does not use nameservice0 but something else for the fs.defaultFs
Digging a little bit, I found that the hadoopConfig
) does not have the properties we set via yarnship/core-site.xml
. Especially, I suspect it is due to the cached HadoopFsFactory
is initialized with an dummy Configuration (code
), which prevents future flinkConfig
getting passed in (code
I am not sure if this is intentional, or has been solved in later releases. Has anyone encountered the same problem? And I would appreciate any suggestions.
"So you have to trust that the dots will somehow connect in your future."