ASF GitHub Bot commented on ROCKETMQ-332:

Jason918 commented on issue #227: [ROCKETMQ-332] fix concurrent bug in 
MappedFileQueue#findMappedFileByOffset, which m…
URL: https://github.com/apache/rocketmq/pull/227#issuecomment-364865146
   @zhouxinyu hi~ Sorry for the delay. This is my fix for this bug in our 
production environment, please check this out. It's been so far so good. :)

This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:

> MappedFileQueue is not thread safe, which will cause message loss.
> ------------------------------------------------------------------
>                 Key: ROCKETMQ-332
>                 URL: https://issues.apache.org/jira/browse/ROCKETMQ-332
>             Project: Apache RocketMQ
>          Issue Type: Bug
>          Components: rocketmq-store
>    Affects Versions: 4.0.0-incubating, 4.1.0-incubating
>            Reporter: Jas0n918
>            Assignee: yukon
>            Priority: Major
>         Attachments: rocketmq.log
> In RocketMQ V3.5.8, there is a readWriteLock in 
> com.alibaba.rocketmq.store.MapedFileQueue, which guarantee thread safety. But 
> in the new org.apache.rocketmq.store.MappedFileQueue, there is not any 
> concurrent control mechanism. 
> when consumer is fetching message(no large lag), broker calls
> org.apache.rocketmq.broker.processor.PullMessageProcessor#processRequest ==>
> org.apache.rocketmq.store.DefaultMessageStore#getMessage  ==>
> org.apache.rocketmq.store.ConsumeQueue#getIndexBuffer ==>
> org.apache.rocketmq.store.MappedFileQueue#findMappedFileByOffset
> but findMappedFileByOffset is not thread safe, as
> org.apache.rocketmq.store.MappedFileQueue#deleteExpiredFile maybe running 
> concurrently(  the size of mappedFiles maybe change) , which will results in 
> ConsumeQueue#getIndexBuffer returns null, causing 
> _nextBeginOffset  = nextOffsetCorrection(offset, 
> consumeQueue.rollNextFile(offset));_+
> which will skip the whole consumeQueue file, any messages left in this 
> ConsumeQueue will not be consumed by client.

This message was sent by Atlassian JIRA

Reply via email to