kafka-jira mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Per Steffensen (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (KAFKA-5716) Connect: When SourceTask.commit it is possible not everthing from SourceTask.poll has been sent
Date Sat, 23 Sep 2017 09:42:00 GMT

    [ https://issues.apache.org/jira/browse/KAFKA-5716?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16177688#comment-16177688
] 

Per Steffensen commented on KAFKA-5716:
---------------------------------------

bq. Yeah, this makes sense. The existing SourceTask.commitRecord(...) method is called after
each source record has been written to Kafka – can you use this to keep track of the offsets
that have been written?

Yeah, that is what I am doing today, because it will calculate something that is closer to
the truth, than just assuming that everything polled has been written. But it is not necessarily
true either. It can happen that (some of) the records that have been polled, but not included
in the offsets-write/flush (which I have demonstrated can happen), has also been sent to outgoing
Kafka AND received acknowledge, leading to the call to {{SourceTask.commitRecord(...)}}. So
at the time of {{SourceTask.commit()}} I claim that it is possible that {{SourceTask.commitRecord(...)}}
has been called for some records not included in the offsets-write/flush.

You suggest several things that I could do, and some of it I am doing. I believe I know very
well, how Kafka and Kafka-Connect works in this area. I can do a lot of things to "workaround"
that fact that I do not really know which offsets has been written/flushed. The thing about
KAFKA-5716 is that I really ought to be told - why not? Or at least the JavaDoc should not
tell me that I am told, when the truth is that I am not :-)

> Connect: When SourceTask.commit it is possible not everthing from SourceTask.poll has
been sent
> -----------------------------------------------------------------------------------------------
>
>                 Key: KAFKA-5716
>                 URL: https://issues.apache.org/jira/browse/KAFKA-5716
>             Project: Kafka
>          Issue Type: Bug
>          Components: KafkaConnect
>            Reporter: Per Steffensen
>            Assignee: Per Steffensen
>            Priority: Minor
>         Attachments: KAFKA-5716.patch
>
>
> Not looking at the very latest code, so the "problem" may have been corrected recently.
If so, I apologize. I found the "problem" by code-inspection alone, so I may be wrong. Have
not had the time to write tests to confirm.
> According to java-doc on SourceTask.commit
> {quote}
> Commit the offsets, up to the offsets that have been returned by \{@link #poll()}. This
> method should block until the commit is complete.
> SourceTasks are not required to implement this functionality; Kafka Connect will record
offsets
> automatically. This hook is provided for systems that also need to store offsets internally
> in their own system.
> {quote}
> As I read this, when commit-method is called, the SourceTask-developer is "told" that
everything returned from poll up until "now" has been sent/stored - both the outgoing messages
and the associated connect-offsets. Looking at the implementation it also seems that this
is what it tries to "guarantee/achieve".
> But as I see read the code, it is not necessarily true
> The following threads are involved
> * Task-thread: WorkerSourceTask has its own thread running WorkerSourceTask.execute.
> * Committer-thread: From time to time SourceTaskOffsetCommitter is scheduled to call
WorkerSourceTask.commitOffsets (from a different thread)
> The two thread synchronize (on the WorkerSourceTask-object) in sendRecord and commitOffsets
respectively, hindering the task-thread to add to outstandingMessages and offsetWriter while
committer-thread is marking what has to be flushed in the offsetWriter and waiting for outstandingMessages
to be empty. This means that the offsets committed will be consistent with what has been sent
out, but not necessarily what has been polled. At least I do not see why the following is
not possible:
> * Task-thread polls something from the task.poll
> * Before task-thread gets to add (all) the polled records to outstandingMessages and
offsetWriter in sendRecords, committer-thread kicks in and does its commiting, while hindering
the task-thread adding the polled records to outstandingMessages and offsetWriter
> * Consistency will not have been compromised, but committer-thread will end up calling
task.commit (via WorkerSourceTask.commitSourceTask), without the records just polled from
task.poll has been sent or corresponding connector-offsets flushed.
> If I am right, I guess there are two way to fix it
> * Either change the java-doc of SourceTask.commit, to something a-la (which I do believe
is true)
> {quote}
> Commit the offsets, up to the offsets that have been returned by \{@link #poll()}
> *and confirmed by a call to \{@link #commitRecord(SourceRecord)}*.
> This method should block until the commit is complete.
> SourceTasks are not required to implement this functionality; Kafka Connect will record
offsets
> automatically. This hook is provided for systems that also need to store offsets internally
> in their own system.
> {quote}
> * or, fix the "problem" so that it actually does what the java-doc says :-)
> If I am not right, of course I apologize for the inconvenience. I would appreciate an
explanation where my code-inspection is not correct, and why it works even though I cannot
see it. I will not expect such an explanation, though.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message