Nicolas Fraison created FLINK-32890:
---------------------------------------

             Summary: Flink app rolled back with old savepoints (3 hours back 
in time) while some checkpoints have been taken in between
                 Key: FLINK-32890
                 URL: https://issues.apache.org/jira/browse/FLINK-32890
             Project: Flink
          Issue Type: Improvement
          Components: Kubernetes Operator
            Reporter: Nicolas Fraison


Here are all details about the issue (based on an app/scenario reproducing the 
issue): * Deployed new release of a flink app with a new operator
 * Flink Operator set the app as stable
 * After some time the app failed and stay in failed state (due to some issue 
with our kafka clusters)
 * Sadly the team in charge of this flink app decided to rollback the app as 
they were thinking it was linked to this new deployment
 * Operator detect: {{Job is not running but HA metadata is available for last 
state restore, ready for upgrade, Deleting JobManager deployment while 
preserving HA metadata.}}  -> rely on last-state (as we do not disable 
fallback), no savepoint taken
 * Flink start JM and deployment of the app. It well find the 3 checkpoints

 * {{Using '/flink-kafka-job-apache-nico/flink-kafka-job-apache-nico' as 
Zookeeper namespace.}}
 * {{Initializing job 'flink-kafka-job' (6b24a364c1905e924a69f3dbff0d26a6).}}
 * {{Recovering checkpoints from 
ZooKeeperStateHandleStore\{namespace='flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobs/6b24a364c1905e924a69f3dbff0d26a6/checkpoints'}.}}
 * {{Found 3 checkpoints in 
ZooKeeperStateHandleStore\{namespace='flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobs/6b24a364c1905e924a69f3dbff0d26a6/checkpoints'}.}}
 * {{Restoring job 6b24a364c1905e924a69f3dbff0d26a6 from Checkpoint 19 @ 
1692268003920 for 6b24a364c1905e924a69f3dbff0d26a6 located at 
}}{{{}s3p://.../flink-kafka-job-apache-nico/checkpoints/6b24a364c1905e924a69f3dbff0d26a6/chk-19{}}}{{{}.{}}}

 * Job failed because of the missing operator

Job 6b24a364c1905e924a69f3dbff0d26a6 reached terminal state FAILED.
org.apache.flink.runtime.client.JobInitializationException: Could not start the 
JobMaster.
Caused by: java.util.concurrent.CompletionException: 
java.lang.IllegalStateException: There is no operator for the state 
f298e8715b4d85e6f965b60e1c848cbe * {{Job 6b24a364c1905e924a69f3dbff0d26a6 has 
been registered for cleanup in the JobResultStore after reaching a terminal 
state.}}
 * {{Clean up the high availability data for job 
6b24a364c1905e924a69f3dbff0d26a6.}}
 * {{Removed job graph 6b24a364c1905e924a69f3dbff0d26a6 from 
ZooKeeperStateHandleStore\{namespace='flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobgraphs'}.}}

 * JobManager restart and try to resubmit the job but the job was already 
submitted so finished

 * {{Received JobGraph submission 'flink-kafka-job' 
(6b24a364c1905e924a69f3dbff0d26a6).}}
 * {{Ignoring JobGraph submission 'flink-kafka-job' 
(6b24a364c1905e924a69f3dbff0d26a6) because the job already reached a 
globally-terminal state (i.e. FAILED, CANCELED, FINISHED) in a previous 
execution.}}
 * {{Application completed SUCCESSFULLY}}

 * Finally the operator rollback the deployment and still indicate that {{Job 
is not running but HA metadata is available for last state restore, ready for 
upgrade}}
 * But the job metadata are not anymore there (clean previously)

(CONNECTED [zookeeper-data-eng-multi-cloud.zookeeper-flink.svc:2181]) /> ls 
/flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobs/6b24a364c1905e924a69f3dbff0d26a6/checkpoints
Path 
/flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobs/6b24a364c1905e924a69f3dbff0d26a6/checkpoints
 doesn't exist
(CONNECTED [zookeeper-data-eng-multi-cloud.zookeeper-flink.svc:2181]) /> ls 
/flink-kafka-job-apache-nico/flink-kafka-job-apache-nico/jobs

(CONNECTED [zookeeper-data-eng-multi-cloud.zookeeper-flink.svc:2181]) /> ls 
/flink-kafka-job-apache-nico/flink-kafka-job-apache-nico
jobgraphs
jobs
leader * The rolled back app from flink operator finally take the last provided 
savepoint as no metadata/checkpoints are available. But this last savepoint is 
an old one as during the upgrade the operator decided to rely on last-state



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to