Tim Armstrong has posted comments on this change. ( http://gerrit.cloudera.org:8080/11317 )
Change subject: IMPALA-7483: abort stuck impalad/catalogd on JVM deadlock ...................................................................... Patch Set 2: (10 comments) http://gerrit.cloudera.org:8080/#/c/11317/2//COMMIT_MSG Commit Message: http://gerrit.cloudera.org:8080/#/c/11317/2//COMMIT_MSG@17 PS2, Line 17: cheap > Just curious if you tried this out on a loaded JVM with a bunch of threads I tried running a local stress test with up to 500 concurrent queries. I added a timer around findDeadlockedThreads() 10/15 times it took 100s of microseconds and 5/15 it took 15-50ms. I'm not sure how valid the measurement is because the time could include the thread being descheduled by the OS and the thread getting stuck in a GC pause. So then I enable JVM debug output with LIBHDFS_OPTS="${LIBHDFS_OPTS} -XX:+PrintGCApplicationStoppedTime -XX:+PrintSafepointStatistics -XX:PrintSafepointStatisticsCount=1" The time spent in the operation was always < 1ms but sometimes it blocked for longer waiting to get into the safepoint. I suspect that's something to do with the sheer number of impalad threads runnning. It looks like there's a periodic safepoint anyway in hotspot every 1s by default so running findDeadlockedThreads() every 60s should only make any problem 1/60 worse. http://gerrit.cloudera.org:8080/#/c/11317/2//COMMIT_MSG@22 PS2, Line 22: findDeadlockedThreads > Maybe worth injecting a deadlock locally and see if findDeadlockedThreads() Extended the testing utility code to induce a deadlock. http://gerrit.cloudera.org:8080/#/c/11317/2//COMMIT_MSG@29 PS2, Line 29: I'm open to ideas about how to write an automated test for this and/or > We could write a Java test that at least invokes the loop that checks for d Yeah I once did an undergrad project where a teammate made heavy use of thread.stop() - it wasn't the most stable bit of code... I think I'll skip automating the test if thats ok. I think we'd want an end-to-end test if anything - a unit test would probably mostly be checking whether findDeadlockThreads() works. We could do a python test that launches a program with a deadlock and checks that it exits, but it feels like overkill. http://gerrit.cloudera.org:8080/#/c/11317/2/be/src/util/jni-util.cc File be/src/util/jni-util.cc: http://gerrit.cloudera.org:8080/#/c/11317/2/be/src/util/jni-util.cc@28 PS2, Line 28: 60 > IMO, configuring it something like 300s is reasonable too since VM seems to My reason for keeping it lower was to reduce the potential duration of an outage resulting from this - if it's 60s then I'd expect the cluster to recover from this in < 2min (up to 60s to detect the hang, then maybe 30s for the statestore to detect the process exit and propagate the failure). Qualitatively a 5 minute+ outage feels worse to me than a shorter outage. As you said though, this is rare so maybe that is ok so long as it resolves itself. http://gerrit.cloudera.org:8080/#/c/11317/2/be/src/util/jni-util.cc@29 PS2, Line 29: "Interval between JVM deadlock checks. If set to 0 or a negative value, deadlock " > nit: mark it (Advanced)? Done http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java File fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java: http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java@176 PS2, Line 176: return; > not related to your change but I just noticed this, can we log this as an e Done http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java@216 PS2, Line 216: for (ThreadInfo thread : deadlockedThreads) LOG.error(thread.toString()); > nit: in theory, given the threads are deadlocked, it should be guaranteed t Done http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java@217 PS2, Line 217: LOG.fatal("Aborting because of deadlocked threads in JVM."); > Should we dump the entire thread stacks (not just the deadlocked ones) and Dumping all the other state feels out of scope for this change. It seems like useful infrastructure but I don't want to do a once-off version of it for this one case. If we want to add a utility to dump all JVM state then we could add a call to that here. http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java@217 PS2, Line 217: LOG.fatal("Aborting because of deadlocked threads in JVM."); > +1 for other thread stacks. Done http://gerrit.cloudera.org:8080/#/c/11317/2/fe/src/main/java/org/apache/impala/util/JvmPauseMonitor.java@217 PS2, Line 217: LOG.fatal("Aborting because of deadlocked threads in JVM."); > oh, also, does log4j "fatal" actually guarantee the process exits? You dont This made use of the fact that we hooked up log4j to glog. Glog fatal does abort the process. That does merit a comment so I added a comment and an exit(1) in case this is used in some context where its not hooked up to glog. -- To view, visit http://gerrit.cloudera.org:8080/11317 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-MessageType: comment Gerrit-Change-Id: If46f879e7ab9eeead0c6e9a09844e1356012898d Gerrit-Change-Number: 11317 Gerrit-PatchSet: 2 Gerrit-Owner: Tim Armstrong <[email protected]> Gerrit-Reviewer: Bharath Vissapragada <[email protected]> Gerrit-Reviewer: Impala Public Jenkins <[email protected]> Gerrit-Reviewer: Philip Zeyliger <[email protected]> Gerrit-Reviewer: Tim Armstrong <[email protected]> Gerrit-Reviewer: Todd Lipcon <[email protected]> Gerrit-Comment-Date: Sat, 25 Aug 2018 00:20:23 +0000 Gerrit-HasComments: Yes
