[
https://issues.apache.org/jira/browse/FLINK-4618?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15489901#comment-15489901
]
Tzu-Li (Gordon) Tai commented on FLINK-4618:
--------------------------------------------
Hi [~melmoth],
Flink achieves exactly-once guarantees by working with offsets that are
checkpointed internally in Flink, not the offsets that are committed back to ZK
/ Kafka. This offset committing is either done periodically or on
checkpointing, depending on the consumer configuration. So, this committed
offset may not always reflect the actual progress of the consumption.
However, on a "fresh" startup of a job ("fresh" startup meaning that the
execution of the job is not an automatic restore from previous failure), the
Kafka consumer respect any existing offsets committed in ZK as starting points.
So, if I am correct, what is actually happening is that, on your second
execution of the job, the Kafka consumer is simply just starting from the
offsets it finds in ZK.
If you want exactly-once for manual job restarts, you would use Flink
savepoints. See
https://ci.apache.org/projects/flink/flink-docs-master/setup/savepoints.html
for more detail.
Otherwise, the exactly-once guarantee refers to job automatic restores across
job failures.
> Last kafka message gets consumed twice when restarting job
> ----------------------------------------------------------
>
> Key: FLINK-4618
> URL: https://issues.apache.org/jira/browse/FLINK-4618
> Project: Flink
> Issue Type: Bug
> Components: Kafka Connector
> Affects Versions: 1.1.2
> Environment: Flink 1.1.2
> Kafka Broker 0.10.0
> Hadoop 2.7.0
> Reporter: Melmoth
>
> There seem to be an issue with the offset management in Flink. When a job is
> stopped and startet again, a message from the previous offset is read again.
> I enabled checkpoints (EXACTLY_ONCE) and FsStateBackend. I started with a new
> consumer group and emitted one record.
> You can cleary see, that the consumer waits for a new record at offset
> 4848911, which is correct. After restarting, it consumes a record at 4848910,
> causing the record to be consumed more than once.
> I checked the offset with the Kafka CMD tools, the commited offset in
> zookeeper is 4848910.
> Here is my log output:
> {code}
> 10:29:24,225 DEBUG org.apache.kafka.clients.NetworkClient
> - Initiating connection to node 2147482646 at hdp1:6667.
> 10:29:24,225 DEBUG
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator - Fetching
> committed offsets for partitions: [myTopic-0]
> 10:29:24,228 DEBUG org.apache.kafka.clients.NetworkClient
> - Completed connection to node 2147482646
> 10:29:24,234 DEBUG
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator - No
> committed offset for partition myTopic-0
> 10:29:24,238 DEBUG org.apache.kafka.clients.consumer.internals.Fetcher
> - Resetting offset for partition myTopic-0 to latest offset.
> 10:29:24,244 DEBUG org.apache.kafka.clients.consumer.internals.Fetcher
> - Fetched offset 4848910 for partition myTopic-0
> 10:29:24,245 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848910
> 10:29:24,773 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848910
> 10:29:25,276 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848910
> -- Inserting a new event here
> 10:30:22,447 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Adding fetched record for partition myTopic-0 with offset 4848910 to
> buffered record list
> 10:30:22,448 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Returning fetched records at offset 4848910 for assigned partition
> myTopic-0 and update position to 4848911
> 10:30:22,451 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:22,953 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:23,456 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:23,887 INFO org.apache.flink.runtime.checkpoint.CheckpointCoordinator
> - Triggering checkpoint 6 @ 1473841823887
> 10:30:23,957 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:23,996 INFO org.apache.flink.runtime.checkpoint.CheckpointCoordinator
> - Completed checkpoint 6 (in 96 ms)
> 10:30:24,196 TRACE
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator - Sending
> offset-commit request with {myTopic-0=OffsetAndMetadata{offset=4848910,
> metadata=''}} to Node(2147482646, hdp1, 6667)
> 10:30:24,204 DEBUG
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator - Committed
> offset 4848910 for partition myTopic-0
> 10:30:24,460 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:24,963 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:30:48,057 INFO org.apache.flink.runtime.blob.BlobServer
> - Stopped BLOB server at 0.0.0.0:2946
> -- Restarting job
> 10:32:01,672 DEBUG org.apache.kafka.clients.NetworkClient
> - Initiating connection to node 2147482646 at hdp1:6667.
> 10:32:01,673 DEBUG
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator - Fetching
> committed offsets for partitions: [myTopic-0]
> 10:32:01,677 DEBUG org.apache.kafka.clients.NetworkClient
> - Completed connection to node 2147482646
> // See below! Shouldn't the offset be 4848911?
> 10:32:01,682 DEBUG org.apache.kafka.clients.consumer.internals.Fetcher
> - Resetting offset for partition myTopic-0 to the committed offset 4848910
> 10:32:01,683 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848910
> 10:32:01,685 DEBUG org.apache.kafka.clients.NetworkClient
> - Initiating connection to node 1001 at hdp1:6667.
> 10:32:01,687 DEBUG org.apache.kafka.clients.NetworkClient
> - Completed connection to node 1001
> // Here record 4848910 gets consumed again!
> 10:32:01,707 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Adding fetched record for partition myTopic-0 with offset 4848910 to
> buffered record list
> 10:32:01,708 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Returning fetched records at offset 4848910 for assigned partition
> myTopic-0 and update position to 4848911
> 10:32:03,721 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:32:04,224 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:32:04,726 TRACE org.apache.kafka.clients.consumer.internals.Fetcher
> - Added fetch request for partition myTopic-0 at offset 4848911
> 10:32:04,894 INFO org.apache.flink.runtime.blob.BlobCache
> - Shutting down BlobCache
> 10:32:04,903 INFO org.apache.flink.runtime.blob.BlobServer
> - Stopped BLOB server at 0.0.0.0:3079
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
