[
https://issues.apache.org/jira/browse/KAFKA-9156?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16984217#comment-16984217
]
Alex Mironov commented on KAFKA-9156:
-------------------------------------
Hey [~lushilin],
Thanks for identifying the issue, I believe your reasoning and research is
correct, as well as the proposed solution. I agree that AtomicReference might
not be enough to protect this code due to "file_=" setter: there could be a
situation when two threads will enter "if (timeIndex.isEmpty)" and "if
(timeIndex.isDefined) else" blocks, thus leading to discrepancy between "_file"
and "timeIndex.get.file" references.
We are running a patch for this in production for a few days now. If confirmed
that it indeed fixes the issue, I'll create an MR.
> LazyTimeIndex & LazyOffsetIndex may cause niobufferoverflow in concurrent
> state
> -------------------------------------------------------------------------------
>
> Key: KAFKA-9156
> URL: https://issues.apache.org/jira/browse/KAFKA-9156
> Project: Kafka
> Issue Type: Bug
> Components: core
> Affects Versions: 2.3.0
> Reporter: shilin Lu
> Priority: Critical
> Attachments: image-2019-11-07-17-42-13-852.png,
> image-2019-11-07-17-44-05-357.png, image-2019-11-07-17-46-53-650.png
>
>
> !image-2019-11-07-17-42-13-852.png!
> this timeindex get function is not thread safe ,may cause create some
> timeindex.
> !image-2019-11-07-17-44-05-357.png!
> When create timeindex not exactly one ,may cause mappedbytebuffer position to
> end. Then write index entry to this mmap file will cause
> java.nio.BufferOverflowException.
>
> !image-2019-11-07-17-46-53-650.png!
>
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
