[ https://issues.apache.org/jira/browse/FLINK-13376?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16891632#comment-16891632 ]
Biao Liu commented on FLINK-13376: ---------------------------------- Currently the implementation of {{ContinuousFileReaderOperator}} is a bit tricky. Under a finite stream scenario, after receiving all splits, it would be blocked in {{close}} to wait the split reader finishing. This behavior is not compatible with {{BoundedOneInput}}. Because {{endInput}} would be triggered before all datum is finished. Before we rewrite the {{ContinuousFileReaderOperator}} through FLIP-27, I propose fixing this in a work-around way. By extending {{ContinuousFileReaderOperator}} with {{BoundedOneInput}}, we could block {{ContinuousFileReaderOperator}} in {{BoundedOneInput}}.{{endInput}}. That way, it keeps the same behavior with the current, and it respects the semantics of {{BoundedOneInput}}. > FileInputFormat can not be used in batch mode of blink-planner > -------------------------------------------------------------- > > Key: FLINK-13376 > URL: https://issues.apache.org/jira/browse/FLINK-13376 > Project: Flink > Issue Type: Sub-task > Components: Connectors / FileSystem, Runtime / Task, Table SQL / > Runtime > Reporter: Jingsong Lee > Priority: Blocker > Fix For: 1.9.0, 1.10.0 > > > FileInputFormat use ContinuousFileReaderOperator in > StreamExecutionEnvironment, and ContinuousFileReaderOperator join the async > thread in close, So before closing, the data is incomplete. > But batch operators use endInput to do some work, this lead to incomplete > records in endInput, and lead to wrong results. > This mean batch mode of blink-planner can not use file related sources... -- This message was sent by Atlassian JIRA (v7.6.14#76016)