voonhous commented on code in PR #19370:
URL: https://github.com/apache/hudi/pull/19370#discussion_r3644482056
##########
hudi-sync/hudi-hive-sync/src/main/java/org/apache/hudi/hive/transaction/lock/HiveMetastoreBasedLockProvider.java:
##########
@@ -225,12 +234,33 @@ private void acquireLockInternal(long time, TimeUnit
unit, LockComponent lockCom
* takes a long time. Must be called only after {@link #lock} has been set.
*/
private void scheduleHeartbeat() {
- Heartbeat heartbeat = new Heartbeat(hiveClient, lock.getLockid());
+ Heartbeat heartbeat = new Heartbeat(hiveClient, lock.getLockid(),
this::onLockLost);
Review Comment:
The lock-id guard in 25abde04 covers the realistic path here: a stale
WAITING tick that fires after the `finally` released the lock now fails the
`lockId` check and gets ignored, and
`releasingALockNormallyIsNotReportedAsLost` pins that behavior down. What
remains is only the pre-release window -- a tick firing in the microseconds
before the `finally` while `lock` still matches, which would also need HMS to
have forgotten a lock it granted moments earlier -- not reachable in practice,
and the next acquire resets the flag anyway.
So no longer a correctness ask. Gating on `lock.getState() ==
LockState.ACQUIRED` would still avoid scheduling a heartbeat the `finally`
immediately cancels, but that is a nit. Fine to resolve as-is.
##########
hudi-sync/hudi-hive-sync/src/test/java/org/apache/hudi/hive/transaction/lock/TestHiveMetastoreBasedLockProviderLockLoss.java:
##########
@@ -0,0 +1,324 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+package org.apache.hudi.hive.transaction.lock;
+
+import org.apache.hudi.common.config.LockConfiguration;
+import org.apache.hudi.common.config.TypedProperties;
+import org.apache.hudi.exception.HoodieLockException;
+
+import org.apache.hadoop.hive.metastore.IMetaStoreClient;
+import org.apache.hadoop.hive.metastore.api.LockComponent;
+import org.apache.hadoop.hive.metastore.api.LockLevel;
+import org.apache.hadoop.hive.metastore.api.LockResponse;
+import org.apache.hadoop.hive.metastore.api.LockState;
+import org.apache.hadoop.hive.metastore.api.LockType;
+import org.apache.hadoop.hive.metastore.api.NoSuchLockException;
+import org.apache.thrift.TException;
+import org.junit.jupiter.api.BeforeEach;
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Field;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.ScheduledFuture;
+import java.util.concurrent.TimeUnit;
+import java.util.function.BooleanSupplier;
+
+import static
org.apache.hudi.common.config.LockConfiguration.HIVE_DATABASE_NAME_PROP_KEY;
+import static
org.apache.hudi.common.config.LockConfiguration.HIVE_TABLE_NAME_PROP_KEY;
+import static
org.apache.hudi.common.config.LockConfiguration.LOCK_HEARTBEAT_INTERVAL_MS_KEY;
+import static org.junit.jupiter.api.Assertions.assertDoesNotThrow;
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNotNull;
+import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertThrows;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.junit.jupiter.api.Assertions.fail;
+import static org.mockito.ArgumentMatchers.any;
+import static org.mockito.ArgumentMatchers.anyLong;
+import static org.mockito.Mockito.doAnswer;
+import static org.mockito.Mockito.doThrow;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.never;
+import static org.mockito.Mockito.times;
+import static org.mockito.Mockito.verify;
+import static org.mockito.Mockito.when;
+
+/**
+ * Unit tests for how {@link HiveMetastoreBasedLockProvider} reacts to the
metastore reporting its
+ * lock as gone, with a mocked {@link IMetaStoreClient} and no live metastore
or ZooKeeper.
+ */
+class TestHiveMetastoreBasedLockProviderLockLoss {
+
+ private static final String DB = "testdb";
+ private static final String TABLE = "testtable";
+ private static final long LOCK_ID = 42L;
+ private static final long OTHER_LOCK_ID = 43L;
+ private static final long HEARTBEAT_INTERVAL_MS = 100L;
+ private static final long AWAIT_TIMEOUT_MS = 30_000L;
+
+ private LockConfiguration lockConfiguration;
+ private LockComponent lockComponent;
+
+ @BeforeEach
+ void setUp() {
+ TypedProperties props = new TypedProperties();
+ props.setProperty(HIVE_DATABASE_NAME_PROP_KEY, DB);
+ props.setProperty(HIVE_TABLE_NAME_PROP_KEY, TABLE);
+ // Keep the ticks short so the scheduled heartbeat fires within the test.
+ props.setProperty(LOCK_HEARTBEAT_INTERVAL_MS_KEY,
String.valueOf(HEARTBEAT_INTERVAL_MS));
+ lockConfiguration = new LockConfiguration(props);
+ lockComponent = new LockComponent(LockType.EXCLUSIVE, LockLevel.TABLE, DB);
+ lockComponent.setTablename(TABLE);
+ }
+
+ @Test
+ void terminalHeartbeatFailureStopsRenewalAndDropsTheLock() throws Exception {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+ when(client.lock(any())).thenReturn(acquiredLock(LOCK_ID));
+ doThrow(new NoSuchLockException("lock " + LOCK_ID + " does not exist"))
+ .when(client).heartbeat(anyLong(), anyLong());
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+
+ // The metastore has expired the lock, so the provider must stop
claiming to hold it.
+ awaitUntil(() -> provider.getLock() == null, "the lost lock must be
dropped by the heartbeat");
+
+ // The heartbeat task latches the failure on its own, so the count below
would stay at one
+ // even if the schedule were left running. Assert the cancellation
itself as well.
+ assertTrue(heartbeatFutureOf(provider).isCancelled(), "the heartbeat
schedule must be cancelled");
+
+ // Give the scheduler several more intervals: no further heartbeat may
be attempted, since
+ // both the schedule and the heartbeat task itself are stopped after a
terminal failure.
+ Thread.sleep(HEARTBEAT_INTERVAL_MS * 5);
+ verify(client, times(1)).heartbeat(0L, LOCK_ID);
+
+ // The writer must learn that it no longer holds exclusivity, and the
provider must not send
+ // a doomed unlock for a lock the metastore has already dropped.
+ assertThrows(HoodieLockException.class, provider::unlock);
+ verify(client, never()).unlock(anyLong());
+ } finally {
+ provider.close();
+ }
+ }
+
+ @Test
+ void transientHeartbeatFailureKeepsRenewingTheLock() throws Exception {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+ when(client.lock(any())).thenReturn(acquiredLock(LOCK_ID));
+ CountDownLatch heartbeats = new CountDownLatch(2);
+ doAnswer(invocation -> {
+ heartbeats.countDown();
+ throw new TException("transient failure");
+ }).when(client).heartbeat(anyLong(), anyLong());
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+
+ assertTrue(heartbeats.await(AWAIT_TIMEOUT_MS, TimeUnit.MILLISECONDS),
+ "a transient failure must not stop the heartbeat schedule");
+ assertNotNull(provider.getLock(), "a transient failure must not drop the
lock");
+
+ provider.unlock();
+ verify(client).unlock(LOCK_ID);
+ assertNull(provider.getLock());
+ } finally {
+ provider.close();
+ }
+ }
+
+ @Test
+ void lockCanBeAcquiredAgainAfterItWasLost() throws Exception {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+ when(client.lock(any())).thenReturn(acquiredLock(LOCK_ID),
acquiredLock(OTHER_LOCK_ID));
+ // Only the first lock is expired by the metastore; heartbeating the
second one succeeds.
+ doThrow(new NoSuchLockException("lock " + LOCK_ID + " does not exist"))
+ .doNothing()
+ .when(client).heartbeat(anyLong(), anyLong());
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+ awaitUntil(() -> provider.getLock() == null, "the lost lock must be
dropped by the heartbeat");
+
+ // Acquiring again must clear the lost-lock state, otherwise the
provider would keep failing
+ // to release locks it holds perfectly well.
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+ provider.unlock();
+
+ verify(client).unlock(OTHER_LOCK_ID);
+ assertNull(provider.getLock());
+ } finally {
+ provider.close();
+ }
+ }
+
+ @Test
+ void lostLockStateDoesNotLeakIntoTheNextAcquire() throws Exception {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+ when(client.lock(any())).thenReturn(acquiredLock(LOCK_ID),
waitingLock(OTHER_LOCK_ID));
+ doThrow(new NoSuchLockException("lock " + LOCK_ID + " does not exist"))
+ .doNothing()
+ .when(client).heartbeat(anyLong(), anyLong());
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+ awaitUntil(() -> provider.getLock() == null, "the lost lock must be
dropped by the heartbeat");
+
+ // The second attempt only got queued, so it leaves no lock behind and
nothing was expired
+ // by the metastore this time round.
+ assertFalse(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+
+ // Releasing must not report the loss that belonged to the previous lock.
+ assertDoesNotThrow(provider::unlock);
+ } finally {
+ provider.close();
+ }
+ }
+
+ @Test
+ void unlockStaysSilentWhenNoLockWasEverHeld() {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ // Releasing a lock that was never acquired is still a no-op: only a
lock the metastore took
+ // away is reported as a failure.
+ assertDoesNotThrow(provider::unlock);
+ } finally {
+ provider.close();
+ }
+ }
+
+ @Test
+ void staleHeartbeatFromAReleasedLockDoesNotDropTheNextLock() throws
Exception {
+ IMetaStoreClient client = mock(IMetaStoreClient.class);
+ when(client.lock(any())).thenReturn(acquiredLock(LOCK_ID),
acquiredLock(OTHER_LOCK_ID));
+ CountDownLatch staleTickStarted = new CountDownLatch(1);
+ CountDownLatch releaseStaleTick = new CountDownLatch(1);
+ CountDownLatch newLockHeartbeats = new CountDownLatch(3);
+ doAnswer(invocation -> {
+ long heartbeatedLockId = invocation.getArgument(1);
+ if (heartbeatedLockId == LOCK_ID) {
+ // Park this tick inside the RPC until the lock it renews has been
released and another one
+ // taken, then fail it the way the metastore fails a heartbeat for a
lock it no longer has.
+ staleTickStarted.countDown();
+ releaseStaleTick.await();
+ throw new NoSuchLockException("lock " + LOCK_ID + " does not exist");
+ }
+ newLockHeartbeats.countDown();
+ return null;
+ }).when(client).heartbeat(anyLong(), anyLong());
+
+ HiveMetastoreBasedLockProvider provider = new
HiveMetastoreBasedLockProvider(lockConfiguration, client);
+ try {
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+ assertTrue(staleTickStarted.await(AWAIT_TIMEOUT_MS,
TimeUnit.MILLISECONDS),
+ "the heartbeat for the first lock must be in flight before it is
released");
+
+ provider.unlock();
+ assertTrue(provider.acquireLock(1000L, TimeUnit.MILLISECONDS,
lockComponent));
+
+ // Only now does the tick scheduled for the first lock come back,
failing because that lock
+ // was released here. Cancelling its schedule could never have stopped
it.
+ releaseStaleTick.countDown();
+
+ assertTrue(newLockHeartbeats.await(AWAIT_TIMEOUT_MS,
TimeUnit.MILLISECONDS),
+ "the stale tick must not stop the renewal of the lock held now,
which would let the "
+ + "metastore expire a perfectly healthy lock mid-commit");
+ assertNotNull(provider.getLock(), "the stale tick must not drop the lock
held now");
+ assertFalse(heartbeatFutureOf(provider).isCancelled(),
+ "the stale tick must not cancel the schedule of the lock held now");
+
+ assertDoesNotThrow(provider::unlock);
+ verify(client).unlock(OTHER_LOCK_ID);
+ } finally {
+ provider.close();
Review Comment:
If an assert above fails before `releaseStaleTick.countDown()` is reached,
the parked tick waits on the latch forever: `close()` only calls
`executor.shutdown()`, which neither interrupts the `await()` nor lets the
non-daemon scheduler thread exit, so a single failed assertion becomes a hung
surefire JVM instead of a red test. Counting down in the finally keeps a
failure a failure:
```suggestion
releaseStaleTick.countDown();
provider.close();
```
Same for `releaseTick` in `releasingALockNormallyIsNotReportedAsLost` below.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
