pratapaditya04 opened a new pull request, #4203:
URL: https://github.com/apache/gobblin/pull/4203

   ## Problem
   
   `GobblinYarnAppLauncher` monitors the submitted YARN application on a 
background, **non-daemon** `ScheduledExecutorService` and returns from 
`launch()` right after submission. For the Azkaban / Grid-Gateway submitter 
path (`AzkabanGobblinYarnAppLauncher.run()` → `launch()`), `run()` therefore 
returns at *submission* time while the launcher-owned non-daemon threads (the 
status monitor, the `ServiceManager`, the YARN client) stay alive for the 
entire job. The submitter JVM never exits on its own, so the orchestrator's 
watchdog force-kills it — surfacing as a spurious cancellation.
   
   Two further issues with the background-monitor design:
   - The monitor's terminal handler calls `stop()` **from the monitor's own 
worker thread**, and `stop()` then calls 
`shutdownExecutorService(applicationStatusMonitor)` — a thread cannot await its 
own executor's termination, so teardown stalls for the full `awaitTermination` 
window.
   - `main()` calls `launch()` then `System.exit(getExitCode())`; with a 
non-blocking `launch()` that exit raced submission.
   
   ## Fix
   
   Drive the application-status poll **synchronously in the calling (`launch()` 
/ `run()`) thread** and remove the background monitor entirely. In attached 
mode `launch()` now loops — reusing the existing 
`ApplicationReportArrivalEvent` / `GetApplicationReportFailureEvent` 
`@Subscribe` handlers via the synchronous `EventBus` — until the application 
reaches a terminal state, or the launcher is otherwise stopped (graceful 
shutdown or lost-AM visibility), then tears down.
   
   Consequences:
   - `launch()` now represents the full application lifecycle (not just 
submission). Once it returns there are no launcher-owned non-daemon threads 
left, so the submitter JVM exits cleanly; this also makes `main()`'s 
`System.exit(getExitCode())` correct.
   - No background executor → no self-shutdown stall.
   - An interrupt (Azkaban `cancel()`) ends the wait and lets the cancel/stop 
path own the terminal status.
   - The loop also exits on `stopped`, so the lost-AM-visibility path (which 
stops the launcher without setting `applicationCompleted`) cannot block forever.
   - **Detached launches are unchanged** — they skip the wait and return right 
after submission, leaving the application running.
   
   The `applicationStatusMonitor` field, its initialization, and its shutdown 
in `stop()` are removed.
   
   ## Compatibility
   
   The only callers of `launch()` are `main()` and 
`AzkabanGobblinYarnAppLauncher.run()`, both of which want blocking behaviour; 
there are no OSS subclasses of `GobblinYarnAppLauncher`. Detached-mode 
behaviour is preserved.
   
   ## Tests
   
   `:gobblin-yarn:compileJava` clean; existing 
`GobblinYarnAppLauncherTerminalGteTest` passes. Adds three blocking-loop unit 
tests: terminal-immediate, poll-until-complete, and a regression guard that the 
loop exits on `stopped` alone (lost-AM-visibility path).
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to