Jose Luis López created HADOOP-19982:
----------------------------------------
Summary: Re-enable TestDFSClientRetries and
TestBalancerWithHANameNodes in GHA
Key: HADOOP-19982
URL: https://issues.apache.org/jira/browse/HADOOP-19982
Project: Hadoop Common
Issue Type: Sub-task
Components: common, hdfs, test
Reporter: Jose Luis López
h3. What
Remove two entries from \{{.github/gha-tests/exclude-tests.txt}}:
* \{{**/org/apache/hadoop/hdfs/TestDFSClientRetries.java}}
*
\{{**/org/apache/hadoop/hdfs/server/balancer/TestBalancerWithHANameNodes.java}}
h3. Why these two
*TestDFSClientRetries* - fixed by HDFS-17972.
{\{testLeaseRenewAndDFSOutputStreamDeadLock}} installed a
{\{DFSClientFaultInjector}} whose \{{delayWhenRenewLeaseTimeout()}} blocked on
an
unbounded \{{testLatch.await()}}, and never restored the previous injector. If
the test died before \{{NameNode.complete()}} counted the latch down, a lease
renewer thread stayed parked inside the injector holding the \{{LeaseRenewer}}
monitor. That hung the rest of the class until Surefire killed the fork, losing
the class's results - exactly the "fails or aborts" the README describes. The
fix bounds the wait at 30s and both counts the latch down and restores the
injector in a \{{finally}} block. All 13 tests in the class already shut their
clusters down in \{{finally}}.
*TestBalancerWithHANameNodes* - fixed by HDFS-17957.
{\{testBalancerWithHANameNodes}} and \{{testBalancerRequestSBNWithHA}} ran on a
60s budget when \{{waitStoragesNoStale}} alone can legitimately consume 60s, so
the deadline could fire before the balancer even started on a loaded agent.
Both are now 300s. The same change removed dead code that pinned the default
NameNode RPC port.
h3. Risk
Both classes run in the \{{hdfs - other}} matrix entry
(\{{-pl :hadoop-hdfs -DexcludedGroups=slow}}); neither carries
\{{@Tag("slow")}},
so one matrix entry exercises both removals. Neither class is excluded anywhere
else in the repository. Since HDFS-17957 set
{\{junit.jupiter.execution.timeout.default}} to 600s for this module, a
regression in either class surfaces as a named test failure rather than a lost
fork, so the cost of getting this wrong is one red test, not one red job.
h3. Acceptance criteria
Five consecutive successful GHA runs with both entries removed, per
{\{.github/gha-tests/README.md}}.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
