RE: Re: Ignite memory leaks in 2.8.0

[Andrey Davydov]

Sorry, It was message to another mail thread where another way of leaks in connection manager is discussed.   >> thread local logic of connection manager is mess +100500 kill it with fire =))))))   Andrey.   От: Taras Ledkov Отправлено: 24 марта 2020 г. в 11:09 Кому: user@ignite.apache.org Тема: Re: Ignite memory leaks in 2.8.0   Hi, Andrey Hmm. There is ConnectionManager#cleanupConnections to close connections of terminated threads (run periodically with default timeout is 2000 ms). So, if the detached connection is recycled after use and returns into #threadConns it should be closed after the owner thread is terminated. take a look at the test: H2ConnectionLeaksSelfTest#testConnectionLeaks But I wrote: thread local logic of connection manager is mess, hard to understand and promising to many troubles. I think we have to change it. On 23.03.2020 23:00, Andrey Davydov wrote: It seems detached connection NEVER become attached to thread other it was born. Because borrow method always return object related to caller thread. I.e. all detached connection borned in joined thread are not collectable forewer.   So possible reproduce scenario: start separate thread. Run in this thread some logic that creates detached connection, finish and join thread. Remove link to thread. Repeat.   пн, 23 мар. 2020 г., 15:49 Taras Ledkov <tled...@gridgain.com>: Hi, Thanks for your investigation. Root cause is clear. What use-case is causing the leak? I've created the issue to remove mess ThreadLocal logic from ConnectionManager. [1] We 've done it in GG Community Edition and it works OK. [1]. https://issues.apache.org/jira/browse/IGNITE-12804 On 21.03.2020 22:50, Andrey Davydov wrote: A simple diagnostic utility I use to detect these problems:   import java.lang.ref.WeakReference; import java.util.ArrayList; import java.util.LinkedList; import java.util.List; import org.apache.ignite.Ignite; import org.apache.ignite.internal.GridComponent; import org.apache.ignite.internal.IgniteKernal; import org.apache.logging.log4j.LogManager; import org.apache.logging.log4j.Logger; public class IgniteWeakRefTracker {     private static final Logger LOGGER = LogManager.getLogger(IgniteWeakRefTracker.class);     private final String clazz;     private final String testName;     private final String name;     private final WeakReference<Ignite> innerRef;     private final List<WeakReference<GridComponent>> componentRefs = new ArrayList<>(128);     private static final LinkedList<IgniteWeakRefTracker> refs = new LinkedList<>();     private IgniteWeakRefTracker(String testName, Ignite ignite) {         this.clazz = ignite.getClass().getCanonicalName();         this.innerRef = new WeakReference<>(ignite);         this.name = ignite.name();         this.testName = testName;         if (ignite instanceof IgniteKernal) {             IgniteKernal ik = (IgniteKernal) ignite;             List<GridComponent> components = ik.context().components();             for (GridComponent c : components) {                 componentRefs.add(new WeakReference<>(c));             }         }     }     public static void register(String testName, Ignite ignite) {         refs.add(new IgniteWeakRefTracker(testName, ignite));     }     public static void trimCollectedRefs() {         List<IgniteWeakRefTracker> toRemove = new ArrayList<>();         for (IgniteWeakRefTracker ref : refs) {             if (ref.isIgniteCollected()) {                 LOGGER.info("Collected ignite: ignite {} from test {}", ref.getIgniteName(), ref.getTestName());                 toRemove.add(ref);                 if (ref.igniteComponentsNonCollectedCount() != 0) {                     throw new IllegalStateException("Non collected components for collected ignite.");                 }             } else {                 LOGGER.warn("Leaked ignite: ignite {} from test {}", ref.getIgniteName(), ref.getTestName());             }         }         refs.removeAll(toRemove);         LOGGER.info("Leaked ignites count:  {}", refs.size());     }     public static int getLeakedSize() {         return refs.size();     }     public boolean isIgniteCollected() {         return innerRef.get() == null;     }     public int igniteComponentsNonCollectedCount() {         int res = 0;         for (WeakReference<GridComponent> cr : componentRefs) {             GridComponent gridComponent = cr.get();             if (gridComponent != null) {                 LOGGER.warn("Uncollected component: {}", gridComponent.getClass().getSimpleName());                 res++;             }         }         return res;     }     public String getClazz() {         return clazz;     }     public String getTestName() {         return testName;     }     public String getIgniteName() {         return name;     } }     On Fri, Mar 20, 2020 at 11:51 PM Andrey Davydov <andrey.davy...@gmail.com> wrote: I found one more way for leak and understand reason:   this     - value: org.apache.ignite.internal.IgniteKernal #1  <- grid     - class: org.apache.ignite.internal.GridKernalContextImpl, value: org.apache.ignite.internal.IgniteKernal #1   <- ctx     - class: org.apache.ignite.internal.processors.timeout.GridTimeoutProcessor, value: org.apache.ignite.internal.GridKernalContextImpl #3    <- this$0     - class: org.apache.ignite.internal.processors.timeout.GridTimeoutProcessor$CancelableTask, value: org.apache.ignite.internal.processors.timeout.GridTimeoutProcessor #1     <- stmtCleanupTask     - class: org.apache.ignite.internal.processors.query.h2.ConnectionManager, value: org.apache.ignite.internal.processors.timeout.GridTimeoutProcessor$CancelableTask #11      <- arg$1     - class: org.apache.ignite.internal.processors.query.h2.ConnectionManager$$Lambda$174, value: org.apache.ignite.internal.processors.query.h2.ConnectionManager #1       <- recycler     - class: org.apache.ignite.internal.processors.query.h2.ThreadLocalObjectPool, value: org.apache.ignite.internal.processors.query.h2.ConnectionManager$$Lambda$174 #1        <- this$0     - class: org.apache.ignite.internal.processors.query.h2.ThreadLocalObjectPool$Reusable, value: org.apache.ignite.internal.processors.query.h2.ThreadLocalObjectPool #1         <- value     - class: java.lang.ThreadLocal$ThreadLocalMap$Entry, value: org.apache.ignite.internal.processors.query.h2.ThreadLocalObjectPool$Reusable #1          <- [411]     - class: java.lang.ThreadLocal$ThreadLocalMap$Entry[], value: java.lang.ThreadLocal$ThreadLocalMap$Entry #35           <- table     - class: java.lang.ThreadLocal$ThreadLocalMap, value: java.lang.ThreadLocal$ThreadLocalMap$Entry[] #25            <- threadLocals (thread object)     - class: java.lang.Thread, value: java.lang.ThreadLocal$ThreadLocalMap #2   Reason:   org.apache.ignite.internal.processors.query.h2.ConnectionManager has some ThreadLocal fields, including connPool, threadConns,  threadConn, detachedConns etc.   ConnectionManager store Lambdas it this thread local storages, so link to ConnectionManager leaks to thread local context.   And seems that method not valid enoght     private void closeConnections() {         threadConns.values().forEach(set -> set.keySet().forEach(U::closeQuiet));         detachedConns.keySet().forEach(U::closeQuiet);         threadConns.clear();         detachedConns.clear();     }   So when Ignition.start() and Ignition.stop()  was from different thread,  caches not cleared properly and starter thread save link to ConnectionManager via ThreadLocal context. And we get one Ignite instance leak every time.   Im sure you run "tens of thousands nodes during every suite run." But majority of runs may be without Indexing, and start and stop node in same thread.   To reproduce leak, start ignite with indexing, save lint to weak reference, and stop it asynchroniouly in other thread, null local link, check weak ref and see heap dump.   Andrey.   От: Andrey Davydov Отправлено: 18 марта 2020 г. в 18:37 Кому: user@ignite.apache.org Тема: Ignite memory leaks in 2.8.0   Hello,   There are at least two way link to IgniteKernal leaks to GC root and makes it unavailable for GC.   The first one:   this     - value: org.apache.ignite.internal.IgniteKernal #1 <- grid     - class: org.apache.ignite.internal.GridKernalContextImpl, value: org.apache.ignite.internal.IgniteKernal #1   <- ctx     - class: org.apache.ignite.internal.processors.query.h2.IgniteH2Indexing, value: org.apache.ignite.internal.GridKernalContextImpl #2    <- this$0     - class: org.apache.ignite.internal.processors.query.h2.IgniteH2Indexing$10, value: org.apache.ignite.internal.processors.query.h2.IgniteH2Indexing #2     <- serializer     - class: org.h2.util.JdbcUtils, value: org.apache.ignite.internal.processors.query.h2.IgniteH2Indexing$10 #1      <- [5395]     - class: java.lang.Object[], value: org.h2.util.JdbcUtils class JdbcUtils       <- elementData     - class: java.util.Vector, value: java.lang.Object[] #37309        <- classes     - class: sun.misc.Launcher$AppClassLoader, value: java.util.Vector #31         <- contextClassLoader (thread object)     - class: java.lang.Thread, value: sun.misc.Launcher$AppClassLoader #1   org.h2.util.JdbcUtils has static field JavaObjectSerializer serializer, which see IgniteKernal via IgniteH2Indexing. It make closed and stopped IgniteKernal non collectable by GC. If some Ignites run in same JVM, JdbcUtils will always use only one, and it can cause some races.   The second way:   this     - value: org.apache.ignite.internal.IgniteKernal #2 <- grid     - class: org.apache.ignite.internal.GridKernalContextImpl, value: org.apache.ignite.internal.IgniteKernal #2   <- ctx     - class: org.apache.ignite.internal.processors.cache.GridCacheContext, value: org.apache.ignite.internal.GridKernalContextImpl #1    <- cctx     - class: org.apache.ignite.internal.processors.cache.distributed.dht.GridDhtCacheEntry, value: org.apache.ignite.internal.processors.cache.GridCacheContext #24     <- parent     - class: org.apache.ignite.internal.processors.cache.GridCacheMvccCandidate, value: org.apache.ignite.internal.processors.cache.distributed.dht.GridDhtCacheEntry #4      <- [0]     - class: java.lang.Object[], value: org.apache.ignite.internal.processors.cache.GridCacheMvccCandidate #1       <- elements     - class: java.util.ArrayDeque, value: java.lang.Object[] #43259        <- value     - class: java.lang.ThreadLocal$ThreadLocalMap$Entry, value: java.util.ArrayDeque #816         <- [119]     - class: java.lang.ThreadLocal$ThreadLocalMap$Entry[], value: java.lang.ThreadLocal$ThreadLocalMap$Entry #51          <- table     - class: java.lang.ThreadLocal$ThreadLocalMap, value: java.lang.ThreadLocal$ThreadLocalMap$Entry[] #21           <- threadLocals (thread object)     - class: java.lang.Thread, value: java.lang.ThreadLocal$ThreadLocalMap #2   Link to IgniteKernal leaks to ThreadLocal variable, so when we start/stop many instances of Ignite in same jvm during testing, we got many stopped “zomby” ignites on ThreadLocal context of main test thread and it cause OutOfMemory after some dozens of tests.   Andrey.     -- Taras Ledkov Mail-To: tled...@gridgain.com -- Taras Ledkov Mail-To: tled...@gridgain.com