[
https://issues.apache.org/jira/browse/FLINK-4027?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15344654#comment-15344654
]
ASF GitHub Bot commented on FLINK-4027:
---------------------------------------
Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/2108#discussion_r68085762
--- Diff:
flink-streaming-connectors/flink-connector-kafka-base/src/main/java/org/apache/flink/streaming/connectors/kafka/FlinkKafkaProducerBase.java
---
@@ -113,6 +123,14 @@
/** Errors encountered in the async producer are stored here */
protected transient volatile Exception asyncException;
+ /**
+ * Number of unacknowledged records.
+ * There is no need to introduce additional locks because invoke() and
snapshotState() are
+ * never called concurrently. So blocking the snapshotting will lock
the invoke() method until all
+ * pending records have been confirmed.
+ */
--- End diff --
I think the fact that `invoke` and `snapshotState` are mutually exclusive
is not important for the semantics of the `pendingRecords` variable. The reason
is that it will only be incremented in `invoke` and decremented in the
`callbacks` of the Kafka producer.
> FlinkKafkaProducer09 sink can lose messages
> -------------------------------------------
>
> Key: FLINK-4027
> URL: https://issues.apache.org/jira/browse/FLINK-4027
> Project: Flink
> Issue Type: Bug
> Components: Kafka Connector
> Affects Versions: 1.0.3
> Reporter: Elias Levy
> Assignee: Robert Metzger
> Priority: Critical
>
> The FlinkKafkaProducer09 sink appears to not offer at-least-once guarantees.
> The producer is publishing messages asynchronously. A callback can record
> publishing errors, which will be raised when detected. But as far as I can
> tell, there is no barrier to wait for async errors from the sink when
> checkpointing or to track the event time of acked messages to inform the
> checkpointing process.
> If a checkpoint occurs while there are pending publish requests, and the
> requests return a failure after the checkpoint occurred, those message will
> be lost as the checkpoint will consider them processed by the sink.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
