KAFKA-14249: Fix flaky TLSv1.3 idle expiry test (Tls13SelectorTest.testCloseOldestConnection)

[Anshul-creator]

What

This PR addresses the flakiness in Tls13SelectorTest.testCloseOldestConnection, which intermittently failed to observe idle connection expiry under TLSv1.3.

Reproduction of flakiness

To reproduce the flaky behavior prior to this fix:

for i in {1..10}; do
  ./gradlew :clients:test \
    --tests 'org.apache.kafka.common.network.Tls13SelectorTest.testCloseOldestConnection' \
    -Pkafka.test.run.flaky=true \
    -PmaxParallelForks=1 -Dorg.gradle.workers.max=1 -Dtest.forkEvery=1 \
    --rerun-tasks --no-daemon \
  || { echo "Failed on run $i"; break; }
done

You can increase the loop count to 20 or more if needed. In my environment, running 10 iterations was enough to reproduce the failure consistently, so I used 10.

Why the flakiness

The base test in SelectorTest assumed that once a connection is established, idle time measurement could begin immediately. This works under PLAINTEXT, but under TLSv1.3:

• Post-handshake records (e.g., NewSessionTicket) often arrive after the handshake completes.
• If the test jumps the mock clock before draining these records, the first poll processes them and updates the channel’s last-activity timestamp to the advanced time.
• As a result, the connection never appears idle, and the test fails.

---

How

This PR makes the TLSv1.3 override robust by:

• Waiting for the channel to be READY instead of just CONNECTED.
• Performing a short selector.poll(50) immediately after READY to drain any post-handshake records at the current time.
• Advancing the mock time beyond the configured idle threshold (parent uses 5s, here we use 10s for margin).
• Using TestUtils.waitForCondition with small polls until the channel is observed as ChannelState.EXPIRED.

These changes ensure idle expiry is deterministically surfaced without altering production logic.

---

Verification of stability

After applying the changes, the same looped command can be used to confirm the test no longer flakes:

for i in {1..20}; do
  ./gradlew :clients:test \
    --tests 'org.apache.kafka.common.network.Tls13SelectorTest.testCloseOldestConnection' \
    -Pkafka.test.run.flaky=true \
    -PmaxParallelForks=1 -Dorg.gradle.workers.max=1 -Dtest.forkEvery=1 \
    --rerun-tasks --no-daemon \
  || { echo "Failed on run $i"; break; }
done

Here the loop count is set to 20 to provide stronger assurance of stability, though fewer iterations may already be sufficient.

Scope

• Test-only change. No production code modified.
• Existing assertions retained (EXPIRED state observed).

JIRA

KAFKA-14249