[ 
https://issues.apache.org/jira/browse/FLINK-41505?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Rion Williams updated FLINK-41505:
----------------------------------
    Description: 
h3. Problem

FLINK-37670 made {{WatermarkAlignmentEvent}} loss tolerant 
({{OperatorEvent#isLossTolerant}}) and has the {{SourceCoordinator}} keep 
sending it after the enumerator has no more splits. 
{{SubtaskGatewayImpl#sendEvent}} accordingly skips the task failover when such 
an event fails with {{TaskNotRunningException}}.

However, the future of that event is still registered in the 
{{IncompleteFuturesTracker}} and completes exceptionally. If a coordinator 
checkpoint is waiting for pending events at that moment, 
{{OperatorCoordinatorHolder#completeCheckpointOnceEventsAreDone}} fails it:
{noformat}
Coordinator checkpoint <id> for coordinator <op> is awaiting 1 pending events
...
org.apache.flink.util.FlinkException: Failing OperatorCoordinator checkpoint 
because some OperatorEvents before this checkpoint barrier were not received by 
the target tasks.
{noformat}
So an event loss that the gateway treats as harmless still fails the 
checkpoint. There are two ways to get there:
 # The delivery fails with {{TaskNotRunningException}}, either on the 
JobManager ({{Execution#sendOperatorEvent}}) or on the TaskManager. The gateway 
skips the failover, but the tracked future fails.
 # The event is still in flight when the task reaches a terminal state. 
{{ExecutionSubtaskAccess}} then fails all pending event futures with a plain 
{{FlinkException("Task is no longer running")}}, which is not recognized as the 
task not running.

h3. When it happens

Watermark alignment is enabled, a source subtask finishes while an alignment 
announcement (every {{updateInterval}}) is in flight to it, and a coordinator 
checkpoint is taken before that delivery fails. The window is when source 
subtasks finish, which is also when coordinator snapshots tend to be taken:
 * streaming with bounded or finishing sources (checkpoints after tasks 
finished): a failed checkpoint, which counts towards the tolerable checkpoint 
failures;
 * batch job recovery: {{DefaultBatchJobRecoveryHandler#onExecutionFinished}} 
snapshots the coordinators (checkpoint id -1) when an execution finishes, so 
the snapshot future fails while the {{ExecutionVertexFinishedEvent}} is written.

We found it through a job that takes a single coordinator snapshot right after 
the sources reach end of input and cannot retry it, so the race failed the 
whole job in about 1 in 5 runs.

h3. Reproduction

Both paths reproduce deterministically in unit tests.

Path 1, in {{OperatorCoordinatorHolderTest}}. It fails on master at the last 
assertion:
{code:java}
@Test
void testCheckpointSucceedsIfLossTolerantEventFindsTaskNotRunning() throws 
Exception {
    CompletableFuture<Acknowledge> eventSendingResult = new 
CompletableFuture<>();
    final EventReceivingTasks tasks =
            
EventReceivingTasks.createForRunningTasksWithRpcResult(eventSendingResult);
    final OperatorCoordinatorHolder holder =
            createCoordinatorHolder(tasks, TestingOperatorCoordinator::new);

    // Send one loss-tolerant event without finishing it.
    CompletableFuture<Acknowledge> sendResult =
            getCoordinator(holder).getSubtaskGateway(0).sendEvent(new 
TestOperatorEvent(0, true));

    // Trigger one checkpoint.
    CompletableFuture<byte[]> checkpointResult = new CompletableFuture<>();
    holder.checkpointCoordinator(1, checkpointResult);
    getCoordinator(holder).getLastTriggeredCheckpoint().complete(new byte[0]);

    // The target task finished before the event reached it.
    eventSendingResult.completeExceptionally(
            new FlinkException(new TaskNotRunningException("test")));

    assertThat(sendResult).isCompletedExceptionally();
    assertThat(checkpointResult).isCompleted(); // fails: completed 
exceptionally
}
{code}
Path 2, in {{OperatorCoordinatorSchedulerTest}}: deploy the tasks with a 
{{TaskExecutorOperatorEventGatew never completes, send a loss-tolerant event, 
trigger a coordinator checkpoint, then switch the task to{{FINISHED}}. The 
checkpoint future completes exceptionally instead of completing.

h3. Proposed fix

Two small changes:
 # In {{SubtaskGatewayImpl#sendEvent}}, track a future for the checkpoint that 
treats a loss-tolerant eveunningException}} as delivered, i.e. the same 
condition that already skips the failover. The futurereturned to the caller 
still completes exceptionally, so 
{{SubtaskGatewayImplTest#optionalEventsIgnoreTaskNotRunning}} is unchanged.
 # In {{ExecutionSubtaskAccess}}, fail the events that are still in flight when 
the task reaches a terminningException}} instead of a plain {{FlinkException}}, 
so that path is recognized as well.

Tests: the two reproductions above, the same check for the batch snapshot id, 
and a test that an event thill fails the checkpoint. I have a patch ready.

  was:
h3. Problem

FLINK-37670 made {{WatermarkAlignmentEvent}} loss tolerant 
({{{}OperatorEvent#isLossTolerant{}}}) and has the {{SourceCoordinator}} keep 
sending it after the enumerator has no more splits. 
{{SubtaskGatewayImpl#sendEvent}} accordingly skips the task failover when such 
an event fails with {{{}TaskNotRunningException{}}}.

However, the future of that event is still registered in the 
{{IncompleteFuturesTracker}} and completes exceptionally. If a coordinator 
checkpoint is waiting for pending events at that moment, 
{{OperatorCoordinatorHolder#completeCheckpointOnceEventsAreDone}} fails it:
{noformat}
Coordinator checkpoint <id> for coordinator <op> is awaiting 1 pending events
...
org.apache.flink.util.FlinkException: Failing OperatorCoordinator checkpoint 
because some OperatorEvents before this checkpoint barrier were not received by 
the target tasks.
{noformat}
So an event loss that the gateway treats as harmless still fails the checkpoint.

The same happens when the event is still in flight while the task reaches a 
terminal state: {{ExecutionSubtaskAccess}} then fails all pending event futures 
with a plain {{FlinkException("Task is no longer running")}}, which is not 
recognized as the task not running.

h3. When it happens

Watermark alignment is enabled, a source subtask finishes while an alignment 
announcement (every {{{}updateInterval{}}}) is in flight to it, and a 
coordinator checkpoint is taken before that delivery fails. The window is when 
source subtasks finish, which is also when coordinator snapshots tend to be 
taken:
 * streaming with bounded or finishing sources (checkpoints after tasks 
finished): a failed checkpoint, which counts towards the tolerable checkpoint 
failures;
 * batch job recovery: {{DefaultBatchJobRecoveryHandler#onExecutionFinished}} 
snapshots the coordinators (checkpoint id -1) when an execution finishes, so 
the snapshot future fails while the {{ExecutionVertexFinishedEvent}} is written.

We found it through a job that takes a single coordinator snapshot right after 
the sources reach end of input and cannot retry it, so the race failed the 
whole job in about 1 in 5 runs.
h3. Reproduction

Deterministic unit test against {{{}OperatorCoordinatorHolderTest{}}}. It fails 
on master at the last assertion:
{code:java}
@Test
void testCheckpointSucceedsIfLossTolerantEventFindsTaskNotRunning() throws 
Exception {
    CompletableFuture<Acknowledge> eventSendingResult = new 
CompletableFuture<>();
    final EventReceivingTasks tasks =
            
EventReceivingTasks.createForRunningTasksWithRpcResult(eventSendingResult);
    final OperatorCoordinatorHolder holder =
            createCoordinatorHolder(tasks, TestingOperatorCoordinator::new);

    // Send one loss-tolerant event without finishing it.
    CompletableFuture<Acknowledge> sendResult =
            getCoordinator(holder).getSubtaskGateway(0).sendEvent(new 
TestOperatorEvent(0, true));

    // Trigger one checkpoint.
    CompletableFuture<byte[]> checkpointResult = new CompletableFuture<>();
    holder.checkpointCoordinator(1, checkpointResult);
    getCoordinator(holder).getLastTriggeredCheckpoint().complete(new byte[0]);

    // The target task finished before the event reached it.
    eventSendingResult.completeExceptionally(
            new FlinkException(new TaskNotRunningException("test")));

    assertThat(sendResult).isCompletedExceptionally();
    assertThat(checkpointResult).isCompleted(); // fails: completed 
exceptionally
}
{code}
h3. Proposed fix

In {{{}SubtaskGatewayImpl#sendEvent{}}}, track a future for the checkpoint that 
treats a loss-tolerant event failing with {{TaskNotRunningException}} as 
delivered, i.e. the same condition that already skips the failover. The future 
returned to the caller still completes exceptionally, so 
{{SubtaskGatewayImplTest#optionalEventsIgnoreTaskNotRunning}} is unchanged. 
Also fail those in-flight events with {{TaskNotRunningException}}. It is a 
small change in {{SubtaskGatewayImpl}} plus the test above; I have a patch 
ready.


> Loss-tolerant OperatorEvent that hits a finished task fails the coordinator 
> checkpoint
> --------------------------------------------------------------------------------------
>
>                 Key: FLINK-41505
>                 URL: https://issues.apache.org/jira/browse/FLINK-41505
>             Project: Flink
>          Issue Type: Bug
>          Components: Runtime / Coordination
>    Affects Versions: 2.3.0, 2.2.1, 2.1.3
>            Reporter: Rion Williams
>            Priority: Major
>             Fix For: 2.5.0
>
>
> h3. Problem
> FLINK-37670 made {{WatermarkAlignmentEvent}} loss tolerant 
> ({{OperatorEvent#isLossTolerant}}) and has the {{SourceCoordinator}} keep 
> sending it after the enumerator has no more splits. 
> {{SubtaskGatewayImpl#sendEvent}} accordingly skips the task failover when 
> such an event fails with {{TaskNotRunningException}}.
> However, the future of that event is still registered in the 
> {{IncompleteFuturesTracker}} and completes exceptionally. If a coordinator 
> checkpoint is waiting for pending events at that moment, 
> {{OperatorCoordinatorHolder#completeCheckpointOnceEventsAreDone}} fails it:
> {noformat}
> Coordinator checkpoint <id> for coordinator <op> is awaiting 1 pending events
> ...
> org.apache.flink.util.FlinkException: Failing OperatorCoordinator checkpoint 
> because some OperatorEvents before this checkpoint barrier were not received 
> by the target tasks.
> {noformat}
> So an event loss that the gateway treats as harmless still fails the 
> checkpoint. There are two ways to get there:
>  # The delivery fails with {{TaskNotRunningException}}, either on the 
> JobManager ({{Execution#sendOperatorEvent}}) or on the TaskManager. The 
> gateway skips the failover, but the tracked future fails.
>  # The event is still in flight when the task reaches a terminal state. 
> {{ExecutionSubtaskAccess}} then fails all pending event futures with a plain 
> {{FlinkException("Task is no longer running")}}, which is not recognized as 
> the task not running.
> h3. When it happens
> Watermark alignment is enabled, a source subtask finishes while an alignment 
> announcement (every {{updateInterval}}) is in flight to it, and a coordinator 
> checkpoint is taken before that delivery fails. The window is when source 
> subtasks finish, which is also when coordinator snapshots tend to be taken:
>  * streaming with bounded or finishing sources (checkpoints after tasks 
> finished): a failed checkpoint, which counts towards the tolerable checkpoint 
> failures;
>  * batch job recovery: {{DefaultBatchJobRecoveryHandler#onExecutionFinished}} 
> snapshots the coordinators (checkpoint id -1) when an execution finishes, so 
> the snapshot future fails while the {{ExecutionVertexFinishedEvent}} is 
> written.
> We found it through a job that takes a single coordinator snapshot right 
> after the sources reach end of input and cannot retry it, so the race failed 
> the whole job in about 1 in 5 runs.
> h3. Reproduction
> Both paths reproduce deterministically in unit tests.
> Path 1, in {{OperatorCoordinatorHolderTest}}. It fails on master at the last 
> assertion:
> {code:java}
> @Test
> void testCheckpointSucceedsIfLossTolerantEventFindsTaskNotRunning() throws 
> Exception {
>     CompletableFuture<Acknowledge> eventSendingResult = new 
> CompletableFuture<>();
>     final EventReceivingTasks tasks =
>             
> EventReceivingTasks.createForRunningTasksWithRpcResult(eventSendingResult);
>     final OperatorCoordinatorHolder holder =
>             createCoordinatorHolder(tasks, TestingOperatorCoordinator::new);
>     // Send one loss-tolerant event without finishing it.
>     CompletableFuture<Acknowledge> sendResult =
>             getCoordinator(holder).getSubtaskGateway(0).sendEvent(new 
> TestOperatorEvent(0, true));
>     // Trigger one checkpoint.
>     CompletableFuture<byte[]> checkpointResult = new CompletableFuture<>();
>     holder.checkpointCoordinator(1, checkpointResult);
>     getCoordinator(holder).getLastTriggeredCheckpoint().complete(new byte[0]);
>     // The target task finished before the event reached it.
>     eventSendingResult.completeExceptionally(
>             new FlinkException(new TaskNotRunningException("test")));
>     assertThat(sendResult).isCompletedExceptionally();
>     assertThat(checkpointResult).isCompleted(); // fails: completed 
> exceptionally
> }
> {code}
> Path 2, in {{OperatorCoordinatorSchedulerTest}}: deploy the tasks with a 
> {{TaskExecutorOperatorEventGatew never completes, send a loss-tolerant event, 
> trigger a coordinator checkpoint, then switch the task to{{FINISHED}}. The 
> checkpoint future completes exceptionally instead of completing.
> h3. Proposed fix
> Two small changes:
>  # In {{SubtaskGatewayImpl#sendEvent}}, track a future for the checkpoint 
> that treats a loss-tolerant eveunningException}} as delivered, i.e. the same 
> condition that already skips the failover. The futurereturned to the caller 
> still completes exceptionally, so 
> {{SubtaskGatewayImplTest#optionalEventsIgnoreTaskNotRunning}} is unchanged.
>  # In {{ExecutionSubtaskAccess}}, fail the events that are still in flight 
> when the task reaches a terminningException}} instead of a plain 
> {{FlinkException}}, so that path is recognized as well.
> Tests: the two reproductions above, the same check for the batch snapshot id, 
> and a test that an event thill fails the checkpoint. I have a patch ready.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to