Santwana Verma created FLINK-41242:
--------------------------------------
Summary: status.error reports every exception of a failed job as
`org.apache.flink.util.SerializedThrowable`
Key: FLINK-41242
URL: https://issues.apache.org/jira/browse/FLINK-41242
Project: Flink
Issue Type: Bug
Components: Kubernetes Operator
Affects Versions: 1.16.1
Reporter: Santwana Verma
When a job fails, the operator fetches the job result and writes its failure
cause to `status.error.` The JobManager sends that cause as a
SerializedThrowable, and
`FlinkResourceExceptionUtils#convertToFlinkResourceException` sets type from
throwable.getClass().getName(). As a result, type is
`org.apache.flink.util.SerializedThrowable` for the top-level error and for
every entry in throwableList, whatever the job actually threw.
For example, a SQL job that fails while casting 'x1' to INT reports:
```
{"type":"org.apache.flink.util.SerializedThrowable",
"message":"org.apache.flink.runtime.JobException: Recovery is suppressed by
FixedDelayRestartBackoffTimeStrategy(...)",
"throwableList":[\{"type":"org.apache.flink.util.SerializedThrowable","message":"..."},
...]}
```
Tools that read status.error therefore cannot tell exceptions apart by type.
The operator's own kubernetes.operator.exception.label.mapper only matches
messages, so it is not affected. The original class name is only available by
parsing it out of message, because SerializedThrowable's message is the
original's toString().
SerializedThrowable already records the original class
(getOriginalErrorClassName()). The proposed fix is to use it for type, both at
the top level and in throwableList, when the throwable is a
SerializedThrowable. message stays as it is, so readers that parse the class
out of it keep working.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
