[
https://issues.apache.org/jira/browse/FLINK-21132?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17274459#comment-17274459
]
Aljoscha Krettek commented on FLINK-21132:
------------------------------------------
I'll try and expand on what I wrote earlier. The TL;DR is that I still think
sending an {{endOfInput()}} to operators on a stop-with-savepoint is a bug.
I differentiate between two different end-of-inputs: 1) _physical_
end-of-input, and 2) _logical/semantical_ end-of-input. The former is signalled
when, for example, a network connection is being shut down. The latter happens
(should happen) when (bounded) sources have no more data to read and this
information propagates through the pipeline. The motivation for introducing
{{endOfInput()}} were things like hash-join in the SQL runner where an operator
would read from the build side until getting an end-of-input, at which point it
would switch over to reading from the probe side. With these use cases in mind
sending an {{endOfInput()}} is a bug. The same is true for sinks, which will do
some bookkeeping based on knowing that all the input data has been read.
I don't see use cases where operators would be interested in being notified of
the physical end-of-input right now. I could be wrong, of course. 😅
> BoundedOneInput.endInput is called when taking synchronous savepoint
> --------------------------------------------------------------------
>
> Key: FLINK-21132
> URL: https://issues.apache.org/jira/browse/FLINK-21132
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Task
> Affects Versions: 1.10.2, 1.10.3, 1.11.3, 1.12.1
> Reporter: Kezhu Wang
> Assignee: Roman Khachatryan
> Priority: Major
>
> [~elkhand](?) reported on project iceberg that {{BoundedOneInput.endInput}}
> was
> [called|https://github.com/apache/iceberg/issues/2033#issuecomment-765864038]
> when [stopping job with
> savepoint|https://github.com/apache/iceberg/issues/2033#issuecomment-765557995].
> I think it is a bug of Flink and was introduced in FLINK-14230. The
> [changes|https://github.com/apache/flink/pull/9854/files#diff-0c5fe245445b932fa83fdaf5c4802dbe671b73e8d76f39058c6eaaaffd9639faL577]
> rely on {{StreamTask.afterInvoke}} and {{OperatorChain.closeOperators}} will
> only be invoked after *end of input*. But that is not true long before after
> [FLIP-34: Terminate/Suspend Job with
> Savepoint|https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=103090212].
> Task could enter state called
> [*finished*|https://github.com/apache/flink/blob/3a8e06cd16480eacbbf0c10f36b8c79a6f741814/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/StreamTask.java#L467]
> after synchronous savepoint, that is an expected job suspension and stopping.
> [~sunhaibotb] [~pnowojski] [~roman_khachatryan] Could you help confirm this ?
> For full context, see
> [apache/iceberg#2033|https://github.com/apache/iceberg/issues/2033]. I have
> pushed branch
> [synchronous-savepoint-conflict-with-bounded-end-input-case|https://github.com/kezhuw/flink/commits/synchronous-savepoint-conflict-with-bounded-end-input-case]
> in my repository. Test case
> {{SavepointITCase.testStopSavepointWithBoundedInput}} failed due to
> {{BoundedOneInput.endInput}} called.
> I am also aware of [FLIP-147: Support Checkpoints After Tasks
> Finished|https://cwiki.apache.org/confluence/x/mw-ZCQ], maybe the three
> should align on what *finished* means exactly. [~kkl0u] [~chesnay]
> [~gaoyunhaii]
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
