[FLINK-11665] Flink fails to remove JobGraph from ZK even though it reports it did

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.5.5, 1.6.4, 1.7.2, 1.8.0
• Fix Version/s: 1.10.0
• Component/s: Runtime / Coordination
• Labels: pull-request-available
• Assignee: Till Rohrmann
• Reporter: Bashar Abdul Jawad
• Created: 19/Feb/19 23:37
• Updated: 31/Jul/20 10:21
• Resolved: 25/Oct/19 15:39

Description

We recently have seen the following issue with Flink 1.5.5:

Given Flink Job ID 1d24cad26843dcebdfca236d5e3ad82a: 

1- A job is activated successfully and the job graph added to ZK:

Added SubmittedJobGraph(1d24cad26843dcebdfca236d5e3ad82a, null) to ZooKeeper.

2- Job is deactivated, Flink reports that the job graph has been successfully removed from ZK and the blob is deleted from the blob server (in this case S3):

Removed job graph 1d24cad26843dcebdfca236d5e3ad82a from ZooKeeper.

3- JM is later restarted, Flink for some reason attempts to recover the job that it reported earlier it has removed from ZK but since the blob has already been deleted the JM goes into a crash loop. The only way to recover it manually is to remove the job graph entry from ZK:

Recovered SubmittedJobGraph(1d24cad26843dcebdfca236d5e3ad82a, null).	

and

org.apache.flink.fs.s3presto.shaded.com.amazonaws.services.s3.model.AmazonS3Exception: The specified key does not exist. (Service: Amazon S3; Status Code: 404; Error Code: NoSuchKey; Request ID: 1BCDFD83FC4546A2), S3 Extended Request ID: OzZtMbihzCm1LKy99s2+rgUMxyll/xYmL6ouMvU2eo30wuDbUmj/DAWoTCs9pNNCLft0FWqbhTo= (Path: s3://blam-state-staging/flink/default/blob/job_1d24cad26843dcebdfca236d5e3ad82a/blob_p-c51b25cc0b20351f6e32a628bb6e674ee48a273e-ccfa96b0fd795502897c73714185dde3)

My question is under what circumstances would this happen? this seems to happen very infrequently but since the consequence is severe (JM crash loop) we'd like to understand how it would happen.

This all seems a little similar to https://issues.apache.org/jira/browse/FLINK-9575 but that issue is reported fixed in Flink 1.5.2 and we are already on Flink 1.5.5

Issue Links

• depends upon: FLINK-11843 Dispatcher fails to recover jobs if leader change happens during JobManagerRunner termination
• is duplicated by: FLINK-11225 Error state of addedJobGraphs when Dispatcher with concurrent revoking and granting leadership
• relates to: FLINK-9575 Potential race condition when removing JobGraph in HA
• links to: GitHub Pull Request #7889
• links to: GitHub Pull Request #9832