YARN-11993. PublicLocalizer thread exits permanently when pending.remove() returns null - #8755

Depends on (has to be merged first):

• HADOOP-19979. Fix four flaky tests that race with work on other threads. #8717 (HADOOP-19979) - HADOOP-19979. Fix four flaky tests that race with work on other threads.
  clears a CI failure: it fixes TestLogAggregationService, which fails in this PR's precommit, still red in the latest run
  mutual: each PR clears a CI failure of the other, so neither goes green alone - merge them back to back

Required by (waits for this one to be merged):

• HADOOP-19972. Upgrade to Jetty 12 without changing the servlet namespace #8704 (HADOOP-19972) - HADOOP-19972. Upgrade to Jetty 12 without changing the servlet namespace
• HADOOP-19979. Fix four flaky tests that race with work on other threads. #8717 (HADOOP-19979) - HADOOP-19979. Fix four flaky tests that race with work on other threads.
  mutual: each PR clears a CI failure of the other, so neither goes green alone - merge them back to back
• HADOOP-19982. Re-enable TestDFSClientRetries and TestBalancerWithHANameNodes in GHA #8725 (HADOOP-19982) - HADOOP-19982. Re-enable TestDFSClientRetries and TestBalancerWithHANameNodes in GHA

Description of PR

In ResourceLocalizationService$PublicLocalizer.run(), when pending.remove(completed) returns null the loop logs "Localized unknown resource" and returns, terminating the Public Localizer thread for the life of the NodeManager. run()'s finally block then shuts the download pool down, so every later public-resource request is rejected — YARN-1800's handler fails each one rather than hanging it, but no public resource can be localized on that node again until the NM restarts. That is the state reported in YARN-9968, which fixed a different branch of this method.

Expected: an unattributable localized resource is skipped and the localizer keeps serving later requests. This PR uses continue, hoisted out of the try.

The branch is defensive rather than reachable today — addResource() puts queue.submit(...) and pending.put(...) in one synchronized (pending) block for exactly this reason. It matters because YARN-11746 proposes shutting the NodeManager down when this thread exits, which would turn one unknown resource into a node-wide outage.

Hoisting the check also makes the assoc dereference in the ExecutionException handler provably safe, clearing NP_NULL_ON_SOME_PATH_EXCEPTION — the only extant SpotBugs warning in this module, which currently costs every NodeManager PR a -1 spotbugs. The exclude entry is therefore deleted rather than repaired; it had been inert since #8568 restored run() without restoring the name YARN-11912 changed to work().

Still leaked, out of scope: the // TODO delete file. Without assoc there is no LocalResourceRequest to fire a failure event for.

How was this patch tested?

New test TestResourceLocalizationService#testPublicLocalizerSurvivesUnknownResource submits a download straight to the completion queue so the Future is never recorded in pending, then asserts the localizer neither dies nor shuts its pool down. Against unmodified trunk it fails with Public Localizer exited after taking an unknown resource; with the fix it passes.

For code changes:

• Does the title or this PR starts with the corresponding JIRA issue id (e.g. 'HADOOP-17799. Your PR title ...')?
• Object storage: have the integration tests been executed and the endpoint declared according to the connector-specific documentation?
• If adding new dependencies to the code, are these dependencies licensed in a way that is compatible for inclusion under ASF 2.0?
• If applicable, have you updated the LICENSE, LICENSE-binary, NOTICE-binary files?

🤖 Generated with Claude Code