Yun Gao created FLINK-24067:
-------------------------------
Summary: CheckpointBarrierHandler may skip the markAlignmentStart
for alignment -with-timeout checkpoint
Key: FLINK-24067
URL: https://issues.apache.org/jira/browse/FLINK-24067
Project: Flink
Issue Type: Bug
Components: Runtime / Checkpointing
Affects Versions: 1.14.0
Reporter: Yun Gao
{code:java}
04:51:11,727 [Flat Map -> Sink: Unnamed (10/12)#1] WARN
org.apache.flink.runtime.taskmanager.Task [] - Flat Map ->
Sink: Unnamed (10/12)#1 (0d965fd3c0de11dc7fb6793435cda8ba) switched from
RUNNING to FAILED with failure cause: java.lang.IllegalStateException:
Alignment time is less than zero({}). Is the time monotonic?
[-9223369873401849363]
at
org.apache.flink.util.Preconditions.checkState(Preconditions.java:215)
at
org.apache.flink.streaming.runtime.io.checkpointing.CheckpointBarrierHandler.markAlignmentEnd(CheckpointBarrierHandler.java:181)
at
org.apache.flink.streaming.runtime.io.checkpointing.CheckpointBarrierHandler.markAlignmentEnd(CheckpointBarrierHandler.java:177)
at
org.apache.flink.streaming.runtime.io.checkpointing.SingleCheckpointBarrierHandler.checkCheckpointAlignedAndTransformState(SingleCheckpointBarrierHandler.java:253)
at
org.apache.flink.streaming.runtime.io.checkpointing.SingleCheckpointBarrierHandler.processBarrier(SingleCheckpointBarrierHandler.java:240)
at
org.apache.flink.streaming.runtime.io.checkpointing.CheckpointedInputGate.handleEvent(CheckpointedInputGate.java:181)
at
org.apache.flink.streaming.runtime.io.checkpointing.CheckpointedInputGate.pollNext(CheckpointedInputGate.java:159)
at
org.apache.flink.streaming.runtime.io.AbstractStreamTaskNetworkInput.emitNext(AbstractStreamTaskNetworkInput.java:110)
at
org.apache.flink.streaming.runtime.io.StreamOneInputProcessor.processInput(StreamOneInputProcessor.java:65)
at
org.apache.flink.streaming.runtime.tasks.StreamTask.processInput(StreamTask.java:490)
{code}
This is caused due to:
1. SingleCheckpointBarrierHandler first receive a BarrierAnnouncement, which
would call checkNewCheckpoint to reset the currentCheckpointId and the
alignedChannels.
2. SingleCheckpointBarrierHandler then received an EndOfPartition, which would
add the channel to the alignedChannels.
3. SingleCheckpointBarrierHandler then received a barrier, which would found
that the alignedChannels is already not empty, then it would skip the
markAlignmentStart.
We might change 3 to judge if this is the first barrier to receive.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
