HIVE-26089: ported TestAsyncPbRpcProxy to junit5, front-loaded JVM warmup and un-ignored a test

[konstantinb]

What changes were proposed in this pull request?

Test-only changes to TestAsyncPbRpcProxy (no production code, no pom changes — JUnit 5 was
already a llap-client test dependency):

• Ported the class to JUnit 5.
• Front-loaded the one-time JVM initialization into a @BeforeAll warmUp() with its own
  60-second @Timeout: Mockito mock generation for the two mocked types, logging setup, and
  classloading of the classes under test.
• Kept the original 5-second per-test budgets as @Timeout(value = 5, threadMode = SEPARATE_THREAD) — SEPARATE_THREAD is required for JUnit 4-equivalent preemption
  (Jupiter's default mode only interrupts, which would not stop a hard hang).
• Un-ignored testSingleInvocationPerNode, disabled since 2022.

This supersedes the approach of #6440 (@Ignore on testMultipleNodes) while restoring the
coverage removed by the 2022 direct-to-master disable
(cc6c02da9).

Why are the changes needed?

Both recorded failures of this class timed out during one-time JVM bootstrap, never in the
code under test: the original HIVE-26089 report died in classloading (ZipFile.getEntry),
and a recent precommit occurrence
(PR 6729, first run)
died inside the first Mockito.mock() call while ByteBuddy generated the mock class. With
surefire's reuseForks=false, every test class starts a cold JVM, so this cost is paid per
class and lands on whichever timed test runs first. The 5s timeouts exist to catch deadlocks
in RequestManager's lock/condition logic (the tests complete in ~0.5s) — they were never
meant to time JVM startup.

The flake reproduces deterministically with no code changes by simulating a starved executor
with a CPU quota — a fresh JVM per run matches reuseForks=false:

# from the repo root, on master (pre-patch)
mvn -pl llap-client test-compile dependency:build-classpath -Dmdep.outputFile=cp.txt -Dmdep.includeScope=test
podman run --rm --cpus=0.15 -v "$PWD:$PWD" -v "$HOME/.m2:$HOME/.m2" docker.io/eclipse-temurin:21-jdk \
  java -cp "$PWD/llap-client/target/test-classes:$PWD/llap-client/target/classes:$(cat cp.txt)" \
  org.junit.runner.JUnitCore org.apache.hadoop.hive.llap.TestAsyncPbRpcProxy

At --cpus=0.15 the unmodified test fails every run (observed 4/4) with the exact CI stack —
TestTimedOutException frozen inside Mockito/ByteBuddy bootstrap at the first mock() call —
and passes from --cpus=0.35 up (docker works the same as podman here).

Splitting the budgets keeps both signals sharp: a 5s per-test timeout still catches a
deadlock, while an init overrun now fails as warmUp() timed out — an environment problem,
clearly distinguished from a test failure.

Does this PR introduce any user-facing change?

No. Test-only.

How was this patch tested?

• CPU-quota measurements (fresh JVM per run): unmodified test FAILS 4/4 at 0.15 CPU; with this
  patch both tests PASS at 0.15 and 0.10; at 0.05 (1/20 of a core) the init budget fails the
  class with warmUp() timed out after 60 seconds and zero test failures — the intended
  broken-executor signal.
• Negative controls: with the warm-up body emptied, both tests fail preemptively at 5s from
  the JUnit timeout thread (proving SEPARATE_THREAD preemption); with a 1-millisecond init
  budget, the class container fails naming warmUp() before any test runs (proving the
  init/test signal separation).
• mvn test -pl llap-client -Dtest=TestAsyncPbRpcProxy: 2/2 green in ~0.5s (confirms surefire
  discovers the Jupiter tests).
• Checkstyle: no new violations.

[konstantinb]

@abstractdog @Aggarwal-Raghav could you take a look? This supersedes the @Ignore approach of #6440 while restoring the disabled test's coverage — the description has a link to the failing precommit run, a deterministic repro, and before/after measurements under CPU starvation. Per the process note on #6440: could a committer trigger a hive-flaky-check run for this branch? Parameters: GITHUB_USER=konstantinb, GITHUB_REPO=hive, BRANCH=HIVE-26089, OPTS=-pl llap-client -Dtest=TestAsyncPbRpcProxy. (I can view the job but don't have build permission.)

[Aggarwal-Raghav]

Thanks @konstantinb for PR. I'm not a committer to this project so I can't trigger the flaky CI but I have given minor comments. Junit5 changes looks good.

NIT: It might be better to explicitly retain the timeout unit in milliseconds. As Hive progressively migrates to JUnit 5 (or eventually JUnit 6), keeping the exact same values?

[Aggarwal-Raghav]

Maybe use import instead of fully pacakge classification? Also, Logger can be defined as class attribute i.e.

private static final Logger LOG = LoggerFactory.getLogger(TestAsyncPbRpcProxy.class);

[Aggarwal-Raghav]

nit: Unused import

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud