Makes sense, I was expecting something like that. +1 for the PR. Even if we are not 100% that will resolve the problem, getting rid of the statics seems like a good idea and beneficial in any case.
On Thu, May 30, 2019 at 12:02 AM Laurent Goujon <[email protected]> wrote: > Unfortunately, not locally, so I ended in crude debugging and code analysis > to understand where the static reloptcluster created for the static > relbuilder was used in a context which would cause issue in a multithreaded > context. Once I saw several rel nodes getting the parent cluster (the > static table scan) to access the metadata query (which cannot be reused > between threads), that's where I assumed this would be the problem. > > On Wed, May 29, 2019 at 2:58 PM Stamatis Zampetakis <[email protected]> > wrote: > > > Thanks a lot Laurent, much appreciated! > > > > BTW, were you able to reproduce the problem? > > > > On Wed, May 29, 2019 at 11:29 PM Laurent Goujon <[email protected]> > > wrote: > > > > > Here's the pull request: https://github.com/apache/calcite/pull/1240 > > > > > > On Wed, May 29, 2019 at 1:58 PM Laurent Goujon <[email protected]> > > wrote: > > > > > > > Looks like most {{RelNode#create()}} access the > > RelOptCluster/RelBuilder > > > > instance from their child, and some then, perform metadata operation, > > > which > > > > would cause CyclicMetadataException for example. I'll create a fixup > > > patch. > > > > > > > > On Wed, May 29, 2019 at 9:00 AM Laurent Goujon <[email protected]> > > > wrote: > > > > > > > >> I actually did the change from using a static relbuilder to a mix of > > > >> static (to create the scans) and per test because of concurrency > > issues > > > :( > > > >> Maybe the test should only been using a per test-case relbuilder > > > instead, > > > >> but it would mean that there's some code relying on the static > > > relbuilder > > > >> (by accessing it from the scan rel nodes) in a non-static context. > > I'll > > > try > > > >> to debug it too (but so far was unlucky reproducing the jenkins > > > problem). > > > >> > > > >> On Wed, May 29, 2019 at 5:02 AM Francis Chuang < > > > [email protected]> > > > >> wrote: > > > >> > > > >>> In Go, there's a built in race detector that can be used when > running > > > >>> tests using `go test`. > > > >>> > > > >>> There's RacerD [1] from Facebook that can detect races in Java. > > Perhaps > > > >>> this is something that can be looked in to to find the race. > > > >>> > > > >>> [1] https://fbinfer.com/docs/racerd.html > > > >>> > > > >>> On 29/05/2019 9:52 pm, Stamatis Zampetakis wrote: > > > >>> > Good insights, Ruben, Danny! > > > >>> > > > > >>> > Assuming that is indeed a concurrency problem, it will be > difficult > > > to > > > >>> > identify since many parts in Calcite are not thread safe. Not > being > > > >>> able to > > > >>> > reproduce the problem makes the things even worse. > > > >>> > > > > >>> > Given that intermittent test failures occur often on Jenkins, how > > > about > > > >>> > creating a new branch dedicated to debugging? > > > >>> > It can be associated with a new Jenkins jobs (that obviously > > doesn't > > > >>> send > > > >>> > mails to everybody when there are failures) that can be launched > > > >>> on-demand > > > >>> > by the person who is looking into the problem. > > > >>> > > > > >>> > On the other hand, we do not really make an effort to have tests > > that > > > >>> can > > > >>> > be executed concurrently so another alternative would be to run > > tests > > > >>> only > > > >>> > sequentially. > > > >>> > I am not very fan of this approach since it will rather hide > > problems > > > >>> than > > > >>> > solve them. > > > >>> > > > > >>> > > > > >>> > > > > >>> > > > > >>> > > > > >>> > > > > >>> > On Wed, May 29, 2019 at 1:02 PM Yuzhao Chen < > [email protected]> > > > >>> wrote: > > > >>> > > > > >>> >> Thanks Ruben for your good analysis. > > > >>> >> > > > >>> >> What I’m confused is that isn’t the static REL_BUILDER more > prone > > to > > > >>> have > > > >>> >> concurrency problems ? And the pushed scans(EMP_SCAN and > > DEPT_SCAN) > > > >>> are all > > > >>> >> nodes(immutable), how could this be a problem ? > > > >>> >> > > > >>> >> Best, > > > >>> >> Danny Chan > > > >>> >> 在 2019年5月29日 +0800 PM5:37，Ruben Q L <[email protected]>，写道： > > > >>> >>> I'm checking the commit [1] and I see something strange in > > > >>> >> RelOptUtilTest. > > > >>> >>> Maybe I'm wrong and it is nothing, but just in case it may > help: > > > >>> >>> > > > >>> >>> With the latest modification, it seems that we have two > > > >>> RelBuilder(s) in > > > >>> >>> place: > > > >>> >>> - A static one that is created ad-hoc on a static block to > > generate > > > >>> the > > > >>> >>> EMP_SCAN and DEPT_SCAN RelNodes [2] > > > >>> >>> - An instance one to be used in the tests, that is initialized > on > > > >>> >>> the @Before public void setUp() method [3] > > > >>> >>> > > > >>> >>> Before this commit, the EMP_SCAN / DEPT_SCAN were only used to > > read > > > >>> their > > > >>> >>> rowTypes to test some join auxiliary methods. But the new > > > >>> >>> tests testPushDownJoinConditions* actually build a plan and > push > > > >>> these > > > >>> >>> scans into the RelBuilder to be tested [4] (which is a > different > > > one > > > >>> than > > > >>> >>> the static RelBuider that created the scans). > > > >>> >>> Maybe this is no problem generally, but it can potentially be > > under > > > >>> >> certain > > > >>> >>> circumstances?, which would explain the randomness of the > issue. > > > >>> >>> Could this explain the exception? > > > >>> >>> > > > >>> >>> [1] > > > >>> >>> > > > >>> >> > > > >>> > > > > > > https://github.com/apache/calcite/commit/82e7d4e760cb203d31956c55e38e0fdd56119d58 > > > >>> >>> > > > >>> >>> [2] > > > >>> >>> > > > >>> >> > > > >>> > > > > > > https://github.com/apache/calcite/blob/ac40d6951bc8c475ca6804be6d878107cc2ebb13/core/src/test/java/org/apache/calcite/plan/RelOptUtilTest.java#L71 > > > >>> >>> [3] > > > >>> >>> > > > >>> >> > > > >>> > > > > > > https://github.com/apache/calcite/blob/ac40d6951bc8c475ca6804be6d878107cc2ebb13/core/src/test/java/org/apache/calcite/plan/RelOptUtilTest.java#L92 > > > >>> >>> [4] > > > >>> >>> > > > >>> >> > > > >>> > > > > > > https://github.com/apache/calcite/blob/ac40d6951bc8c475ca6804be6d878107cc2ebb13/core/src/test/java/org/apache/calcite/plan/RelOptUtilTest.java#L292 > > > >>> >>> > > > >>> >>> > > > >>> >>> > > > >>> >>> Le mer. 29 mai 2019 à 02:20, Julian Hyde <[email protected]> a > > > écrit > > > >>> : > > > >>> >>> > > > >>> >>>> It’s a tough call. It is probable that the problem existed > > already > > > >>> and > > > >>> >> the > > > >>> >>>> change merely surfaced it. > > > >>> >>>> > > > >>> >>>>> On May 28, 2019, at 5:17 PM, Stamatis Zampetakis < > > > >>> [email protected]> > > > >>> >>>> wrote: > > > >>> >>>>> > > > >>> >>>>> It is not the only test that is failing after commit [1] but > > all > > > >>> the > > > >>> >> new > > > >>> >>>>> tests that were added. > > > >>> >>>>> > > > >>> >>>>> I've seen the problem on Jenkins on all JDKS but I cannot > > > reproduce > > > >>> >> it > > > >>> >>>>> locally. > > > >>> >>>>> I guess we have to do with a race condition most likely due > to > > > the > > > >>> >>>>> concurrent execution of tests with surefire. > > > >>> >>>>> > > > >>> >>>>> Should we revert the commit till we find a solution? > > > >>> >>>>> > > > >>> >>>>> [1] > > > >>> >>>>> > > > >>> >>>> > > > >>> >> > > > >>> > > > > > > https://github.com/apache/calcite/commit/82e7d4e760cb203d31956c55e38e0fdd56119d58 > > > >>> >>>>> > > > >>> >>>>> On Tue, May 28, 2019 at 7:57 PM Julian Hyde < > [email protected]> > > > >>> >> wrote: > > > >>> >>>>> > > > >>> >>>>>> I have seen this intermittent failure 3 times in the last > > week: > > > >>> >>>>>> > > > >>> >>>>>> [INFO] Running org.apache.calcite.plan.RelOptUtilTest > > > >>> >>>>>> [ERROR] Tests run: 11, Failures: 0, Errors: 1, Skipped: 0, > > Time > > > >>> >> elapsed: > > > >>> >>>>>> 0.411 s <<< FAILURE! - in > > org.apache.calcite.plan.RelOptUtilTest > > > >>> >>>>>> [ERROR] > > > >>> >>>>>> > > > >>> >>>> > > > >>> >> > > > >>> > > > > > > testPushDownJoinConditionsWithExpandedIsNotDistinctUsingCase(org.apache.calcite.plan.RelOptUtilTest) > > > >>> >>>>>> Time elapsed: 0.349 s <<< ERROR! > > > >>> >>>>>> org.apache.calcite.rel.metadata.CyclicMetadataException > > > >>> >>>>>> at > > > >>> >>>>>> > > > >>> >>>> > > > >>> >> > > > >>> > > > > > > org.apache.calcite.plan.RelOptUtilTest.testPushDownJoinConditionsWithExpandedIsNotDistinctUsingCase(RelOptUtilTest.java:445) > > > >>> >>>>>> > > > >>> >>>>>> I have seen it on Oracle JDK 12 and OpenJDK 10. The test was > > > only > > > >>> >> added > > > >>> >>>> on > > > >>> >>>>>> May 22 so I assume that it will continue to fail > > intermittently > > > >>> >> until > > > >>> >>>> we do > > > >>> >>>>>> something. > > > >>> >>>>>> > > > >>> >>>>>> Anyone have any ideas? > > > >>> >>>>>> > > > >>> >>>>>> Laurent, As you added the test can you please look into it? > > > >>> >>>>>> > > > >>> >>>>>> Julian > > > >>> >>>>>> > > > >>> >>>>>> > > > >>> >>>> > > > >>> >>>> > > > >>> >> > > > >>> > > > > >>> > > > >> > > > > > >
