[GOBBLIN-2011] Fix bug where another host could report the job as skipped, then the …

[Will-Lo]

…skipped jobstatus is used to check for concurrent flows

Dear Gobblin maintainers,

Please accept this PR. I understand that it will not be reviewed until I have checked off all the steps below!

JIRA

• My PR addresses the following Gobblin JIRA issues and references them in the PR title. For example, "[GOBBLIN-XXX] My Gobblin PR"
  • https://issues.apache.org/jira/browse/GOBBLIN-2011

Description

• Here are some details about my PR, including screenshots (if applicable):
  There's a bug that causes GaaS multileader to kick off unintended concurrent flows which happens in the order described below:

1. Host A checks the latest flow execution status to ensure the prior flow is not running, sees that the prior execution is still running.
2. Host A fails the flow pending execution as it cannot run concurrent flow, this emits a FAILED event to GaaS which is ingested by the JobStatusMonitor.
3. Host B checks the latest flow execution status, sees the current flow execution ID which is FAILED (considered a finished flow).
4. Host B kicks off the pending flow execution when it shouldn't be.

To resolve this, we need to ensure that we are looking at the past 2 flow executions, and follow the behavior:

1. If there is no prior execution, kick off the pending flow
2. If the prior execution is IN PROGRESS, we want to indicate that there is a concurrent flow and block the pending execution.
3. If the prior execution is FINISHED, then we want to kick off the pending execution (rely on the DagManager for deduplication of flows because we do not know if the host managing this pending flow is running behind the other hosts).

Tests

• My PR adds the following unit tests OR does not need testing for this extremely good reason:
  Unit tests

Commits

• My commits all reference JIRA issues in their subject lines, and I have squashed multiple commits if they address the same issue. In addition, my commits follow the guidelines from "How to write a good git commit message":
  1. Subject is separated from body by a blank line
  2. Subject is limited to 50 characters
  3. Subject does not end with a period
  4. Subject uses the imperative mood ("add", not "adding")
  5. Body wraps at 72 characters
  6. Body explains "what" and "why", not "how"

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 62.50000% with 3 lines in your changes are missing coverage. Please review.

Project coverage is 46.66%. Comparing base (a78ee54) to head (74360c4).
Report is 1 commits behind head on master.

Files	Patch %	Lines
...obblin/service/monitoring/FlowStatusGenerator.java	62.50%	0 Missing and 3 partials ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##             master    #3888      +/-   ##
============================================
- Coverage     46.67%   46.66%   -0.01%     
+ Complexity    11154    11152       -2     
============================================
  Files          2219     2219              
  Lines         87657    87660       +3     
  Branches       9621     9624       +3     
============================================
- Hits          40911    40910       -1     
- Misses        43055    43057       +2     
- Partials       3691     3693       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[umustafi]

can you make a comment to make clear that the first one is the most recent and may or may not match the pending flowExecutionId attempt and the second index is an older one? It's a bit hard to keep track of what's going on here without reading ur commit desc

[arjun4084346]

I think we should not set the status "Failed" when the last execution is running. We should instead emit a new event "SKIPPED". With this any further execution should be able to correctly decide whether a new job should run or not.
This will make the code more maintainable.

[Will-Lo]

@arjun4084346 I agree that we should emit a new event SKIPPED but we will need to make sure that this is compatible with all clients, which can cause issues if they are set to expect only certain outputs.