[
https://issues.apache.org/jira/browse/KUDU-3755?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18070661#comment-18070661
]
ASF subversion and git services commented on KUDU-3755:
-------------------------------------------------------
Commit 99d498fb5fe30828a2f52b4b6b8609f2c97ddaeb in kudu's branch
refs/heads/master from Marton Greber
[ https://gitbox.apache.org/repos/asf?p=kudu.git;h=99d498fb5 ]
KUDU-3755 tool: add leave_files to control shell
Add a leave_files boolean field to CreateClusterRequestPB so callers
can opt out of the unconditional cluster-root deletion that the
control shell performs on exit.
Add TestLeaveFiles, TestDeleteOnExit, and
TestPreserveFilesOnAbnormalExit to ControlShellToolTest to verify both
normal exit paths and that files are preserved when the shell exits
abnormally.
Change-Id: Ifa09d63d614190d25cb9b1c6706fd0db992eb5eb
Reviewed-on: http://gerrit.cloudera.org:8080/24122
Reviewed-by: Alexey Serbin <[email protected]>
Tested-by: Alexey Serbin <[email protected]>
> Python/Java test scaffolding: preserve cluster data on test failure
> -------------------------------------------------------------------
>
> Key: KUDU-3755
> URL: https://issues.apache.org/jira/browse/KUDU-3755
> Project: Kudu
> Issue Type: Bug
> Reporter: Marton Greber
> Priority: Minor
>
> The C++ test scaffolding preserves the cluster data directory when a test
> fails, controlled by --test_leave_files=on_failure (the default). See
> src/kudu/util/test_util.cc:
> {code:java}
> KuduTest::~KuduTest() {
> // Reset the flags first to prevent them from affecting test directory
> cleanup.
> flag_saver_.reset(); // Clean up the test directory in the destructor
> instead of a TearDown
> // method. This is better because it ensures that the child-class
> // dtor runs first -- so, if the child class is using a minicluster, etc,
> // we will shut that down before we remove files underneath.
> if (FLAGS_test_leave_files == "always") {
> LOG(INFO) << "-----------------------------------------------";
> LOG(INFO) << "--test_leave_files specified, leaving files in " <<
> test_dir_;
> } else if (FLAGS_test_leave_files == "on_failure" && HasFailure()) {
> LOG(INFO) << "-----------------------------------------------";
> LOG(INFO) << "Had failures, leaving test files at " << test_dir_;
> } else {
> VLOG(1) << "Cleaning up temporary test files...";
> WARN_NOT_OK(env_->DeleteRecursively(test_dir_),
> "Couldn't remove test files");
> }
> }{code}
> The Python and Java test scaffolding do not follow this convention. Both
> delegate cluster data cleanup entirely to the kudu test mini_cluster control
> shell, which unconditionally deletes the cluster root on exit regardless of
> test outcome (src/kudu/tools/tool_action_test.cc):
> {code:java}
> // Normal exit, clean up cluster root.
> if (cluster) {
> cluster->Shutdown();
> WARN_NOT_OK(Env::Default()->DeleteRecursively(cluster->cluster_root()),
> "Could not delete cluster root");
> }
> return Status::OK(); {code}
> Java explicitly documents this delegation
> (java/kudu-test-utils/src/main/java/org/apache/kudu/test/cluster/MiniKuduCluster.java):
> {code:java}
> // If a cluster root was not set, create a unique temp directory to use.
> // The mini cluster will clean this directory up on exit.
> try {
> File tempRoot = TempDirUtils.makeTempDirectory("mini-kudu-cluster",
> TempDirUtils.DeleteOnExit.NO_DELETE_ON_EXIT); {code}
> What needs to be done:
> # Add a leave_files boolean field to CreateClusterRequestPB in
> src/kudu/tools/tool.proto so callers can opt out of the unconditional cleanup
> in the control shell
> # Implement the opt-out in tool_action_test.cc
> # Python (python/kudu/tests/common.py): detect test failure (Python 3:
> self._outcome.errors; Python 2: result.failures), pass leaveFiles: true to
> the control shell, and manage directory lifetime based on a
> KUDU_TEST_LEAVE_FILES env var (on_failure/always/never) mirroring the C++ flag
> # Java (KuduTestHarness.java / MiniKuduCluster.java): same, using JUnit 4's
> TestWatcher.failed()/succeeded() callbacks to detect outcome
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
