[
https://issues.apache.org/jira/browse/HBASE-9490?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13762830#comment-13762830
]
Nicolas Liochon commented on HBASE-9490:
----------------------------------------
Yes. We should try to have as much "real unit tests" as possible. A real unit
tests being something that can work in parallel with the other tests. That's a
continuous effort, difficult to do, but necessary if we want to limit our test
time. Agreed, there is not much individual & short term incentive
Surefire now supports reuse fork, this would allow to parallelize the small
tests at no cost. Surefire 2.15 was not working with HBase, but the 2.16 is now
available. If you can give it a try it would be great (see the old patch in
HBASE-4955).
-PlocalTests should be used only to run a subset of tests. Even on your local
machine, you should run the tests in parallel (and usually your dev machine
will be more powerful than the apache build machine). This is explained in the
hbase book, 16.7.3.5. Running tests faster).
> Provide independent execution environment for small tests
> ---------------------------------------------------------
>
> Key: HBASE-9490
> URL: https://issues.apache.org/jira/browse/HBASE-9490
> Project: HBase
> Issue Type: Improvement
> Reporter: Vasu Mariyala
> Assignee: Vasu Mariyala
> Attachments: 0.94-Independent-Test-Execution.patch,
> 0.96-trunk-Independent-Test-Execution.patch
>
>
> Some of the state related to schema metrics is stored in static variables and
> since the small test cases are run in a single jvm, it is causing random
> behavior in the output of the tests.
> An example scenario is the test case failures in HBASE-8930
> {code}
> for (SchemaMetrics cfm : tableAndFamilyToMetrics.values()) {
> if (metricName.startsWith(CF_PREFIX + CF_PREFIX)) {
> throw new AssertionError("Column family prefix used twice: " +
> metricName);
> }
> {code}
> The above code throws an error when the metric name starts with "cf.cf.". It
> would be helpful if any one sheds some light on the reason behind checking
> for "cf.cf."
> The scenarios in which we would have a metric name start with "cf.cf." are as
> follows (See generateSchemaMetricsPrefix method of SchemaMetrics)
> a) The column family name should be "cf"
> AND
> b) The table name is either "" or use table name globally should be false
> (useTableNameGlobally variable of SchemaMetrics).
> Table name is empty only in the case of ALL_SCHEMA_METRICS which has the
> column family as "". So we could rule out the
> possibility of the table name being empty.
> Also to note, the variables "useTableNameGlobally" and
> "tableAndFamilyToMetrics" of SchemaMetrics are static and are shared across
> all the tests that run in a single jvm. In our case, the profile runAllTests
> has the below configuration
> {code}
> <surefire.firstPartForkMode>once</surefire.firstPartForkMode>
> <surefire.firstPartParallel>none</surefire.firstPartParallel>
> <surefire.firstPartThreadCount>1</surefire.firstPartThreadCount>
>
> <surefire.firstPartGroups>org.apache.hadoop.hbase.SmallTests</surefire.firstPartGroups>
> {code}
> Hence all of our small tests run in a single jvm and share the above
> variables "useTableNameGlobally" and "tableAndFamilyToMetrics".
> The reasons why the order of execution of the tests caused this failure are
> as follows
> a) A bunch of small tests like TestMemStore, TestSchemaConfiguredset set the
> useTableNameGlobally to false. But these tests don't create tables that have
> the column family name as "cf".
> b) If the tests in step (a) run before the tests which create table/regions
> with column family 'cf', metric names would start with "cf.cf."
> c) If any of other tests, like the failed tests(TestScannerSelectionUsingTTL,
> TestHFileReaderV1, TestScannerSelectionUsingKeyRange), validate schema
> metrics, they would fail as the metric names start with "cf.cf."
> On my local machine, I have tried to re-create the failure scenario by
> changing the sure fire test configuration and creating a simple (TestSimple)
> which just creates a region for the table 'testtable' and column family 'cf'.
> {code}
> TestSimple.java
> ------------------------------------------------------------------
> @Before
> public void setUp() throws Exception {
> HTableDescriptor htd = new HTableDescriptor(TABLE_NAME_BYTES);
> htd.addFamily(new HColumnDescriptor(FAMILY_NAME_BYTES));
> HRegionInfo info = new HRegionInfo(TABLE_NAME_BYTES, null, null, false);
> this.region = HRegion.createHRegion(info, TEST_UTIL.getDataTestDir(),
> TEST_UTIL.getConfiguration(), htd);
> Put put = new Put(ROW_BYTES);
> for (int i = 0; i < 10; i += 2) {
> // puts 0, 2, 4, 6 and 8
> put.add(FAMILY_NAME_BYTES, Bytes.toBytes(QUALIFIER_PREFIX + i), i,
> Bytes.toBytes(VALUE_PREFIX + i));
> }
> this.region.put(put);
> this.region.flushcache();
> }
> @Test
> public void testFilterInvocation() throws Exception {
> System.out.println("testing");
> }
> @After
> public void tearDown() throws Exception {
> HLog hlog = region.getLog();
> region.close();
> hlog.closeAndDelete();
> }
> Successful run:
> -------------------------------------------------------
> T E S T S
> -------------------------------------------------------
> 2013-09-09 15:38:03.478 java[46562:db03] Unable to load realm mapping info
> from SCDynamicStore
> Running org.apache.hadoop.hbase.filter.TestSimple
> Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.342 sec
> Running org.apache.hadoop.hbase.io.hfile.TestHFileReaderV1
> Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.085 sec
> Running org.apache.hadoop.hbase.io.hfile.TestScannerSelectionUsingKeyRange
> Tests run: 3, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.217 sec
> Running org.apache.hadoop.hbase.io.hfile.TestScannerSelectionUsingTTL
> Tests run: 6, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 12.618 sec
> Running org.apache.hadoop.hbase.regionserver.TestMemStore
> Tests run: 24, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 26.542 sec
> Results :
> Tests run: 35, Failures: 0, Errors: 0, Skipped: 0
> ------------------------------------------------------------------
> Failed run order:
> -------------------------------------------------------
> T E S T S
> -------------------------------------------------------
> 2013-09-09 15:43:21.466 java[46890:db03] Unable to load realm mapping info
> from SCDynamicStore
> Running org.apache.hadoop.hbase.regionserver.TestMemStore
> Tests run: 24, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 11.967 sec
> Running org.apache.hadoop.hbase.io.hfile.TestScannerSelectionUsingTTL
> Tests run: 6, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 12.659 sec
> Running org.apache.hadoop.hbase.io.hfile.TestScannerSelectionUsingKeyRange
> Tests run: 3, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.15 sec
> Running org.apache.hadoop.hbase.filter.TestSimple
> Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.031 sec
> Running org.apache.hadoop.hbase.io.hfile.TestHFileReaderV1
> Tests run: 1, Failures: 1, Errors: 0, Skipped: 0, Time elapsed: 24.883 sec
> <<< FAILURE!
> Results :
> Failed tests:
> testReadingExistingVersion1HFile(org.apache.hadoop.hbase.io.hfile.TestHFileReaderV1):
> Column family prefix used twice: cf.cf.bt.Data.fsReadnumops
> Tests run: 35, Failures: 1, Errors: 0, Skipped: 0
> [INFO]
> ------------------------------------------------------------------------
> [INFO] BUILD FAILURE
> [INFO]
> ------------------------------------------------------------------------
> {code}
> In the failed scenario, the below has happened
> a) TestMemStore sets the useTableNameGlobally to false
> b) TestScannerSelectionUsingKeyRange, TestScannerSelectionUsingKeyRange are
> successful as they don't create table with column family name "cf"
> c) TestSimple creates a region for table 'testtable' and column family 'cf'.
> Since useTableNameGlobally is set to false, it would create metric names that
> start with "cf.cf."
> d) TestHFileReaderV1 while validating metrics would fail as the metric names
> start with "cf.cf."
> The reason why this has been exposed due to this patch is because TestSimple
> is TestInvocationRecordFilter. The executions of the build 1136, 1137 and
> 1138 which have been executed after this patch have a different order of
> executions when compared to the failed builds 1139 & 1140.
> One simple fix to address the issue would have been to change the column
> family name from "cf" to "mycf" in the TestInvocationRecordFilter. But to
> avoid future occurrences of these issues, I would suggest setting the
> "surefire.firstPartForkMode" to "always" similar to the settings we use while
> running localTests, medium & large tests.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
