[ https://issues.apache.org/jira/browse/FLINK-33466?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17783896#comment-17783896 ]
Jonas Weile edited comment on FLINK-33466 at 11/8/23 6:49 AM: -------------------------------------------------------------- See the following issues for the same problem: FLINK-31006 FLINK-29674 Apache Kafka Connector‘s “ setBounded” not valid - ASF Jira It seems, that the two possible solutions are as follows: 1. Either the sources must signal the NoMoreSplitsEvent again after restoring from a snapshot or 2. the SourceReaderBase should snapshot the noMoreSplitsAssignment variable. Which approach would be most viable? was (Author: JIRAUSER302117): See the following issues for the same problem: [FLINK-31006] job is not finished when using pipeline mode to run bounded source like kafka/pulsar - ASF JIRA (apache.org) [FLINK-29674] Apache Kafka Connector‘s “ setBounded” not valid - ASF Jira It seems, that the two possible solutions are as follows: 1. Either the sources must signal the NoMoreSplitsEvent again after restoring from a snapshot or 2. the SourceReaderBase should snapshot the noMoreSplitsAssignment variable. Which approach would be most viable? > Bounded Kafka source never finishes after restore from savepoint > ---------------------------------------------------------------- > > Key: FLINK-33466 > URL: https://issues.apache.org/jira/browse/FLINK-33466 > Project: Flink > Issue Type: Bug > Components: Connectors / Common > Affects Versions: 1.17.1 > Reporter: Jonas Weile > Priority: Major > > When setting up a bounded Kafka source, if the job is restored from a > savepoint before the source has finished, then the Kafka source will never > transition to a finished state. > This seems to be because the noMoreSplitsAssignment variable in the > SourceReaderBase class is not snapshotted. Therefore, after restoring from a > checkpoint/savepoint, the noMoreSplitsAssignment variable will default to > false, and the first condition in the private helper method > finishedOrAvailableLater() in the SourceReaderBase class will always evaluate > to true. > Since this originates in the base class, the problem should hold for all > source types, not just kafka. > > Would it make sense to snapshot the noMoreSplitsAssigntments variable? > I would love to take this on as a first task if appropriate. -- This message was sent by Atlassian Jira (v8.20.10#820010)