[jira] [Commented] (IOTDB-5093) Fix compaction selector bug

["周沛辰 (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/IOTDB-5093?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17642504#comment-17642504 ] 

周沛辰 commented on IOTDB-5093:
----------------------------

*跨空间选文件流程：* # 
从TsFileManager里获取seqFiles和unseqFiles

 # 
过滤掉被删除的顺序文件和被删除、isCompacting、isCandidate的乱序文件，放入CandidateList

 # 
从CandidateList里选择有重叠的顺乱序文件，每选一个乱序文件，就检查其顺序文件是否有效（若isCompacting、isCandidate、isDeleted则无效），若否则返回

*Bug产生：*
!image-2022-12-02-21-29-16-925.png|width=853,height=205!
文件3 4正在被其他线程合并，接着来了跨空间选文件，选文件流程： # 
获取的seqFiles里有1 2 3 4 5。（接着其他合并线程对文件3'与3 4进行内存置换，然后 3 4 被置为deleted并删除了）

 # 
第二步的时候就会过滤掉 3 4，将 1 2 5放入CandidateList里

 # 
从CandidateList里进行选文件，选中1 2 5 6，导致合出来的目标文件与其他文件有重叠

*解决方法：* # 
获取 1 2 3 4 5

 # 
不对顺序文件进行过滤，因此顺序CandidateList里有1 2 3 4 5

 # 
选文件发现3 4 正在被其他合并占用或者被删除 丢弃文件

> Fix compaction selector bug
> ---------------------------
>
>                 Key: IOTDB-5093
>                 URL: https://issues.apache.org/jira/browse/IOTDB-5093
>             Project: Apache IoTDB
>          Issue Type: Bug
>    Affects Versions: master branch, 0.13.0-SNAPSHOT
>            Reporter: 周沛辰
>            Assignee: 周沛辰
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 1.0.0
>
>         Attachments: image-2022-12-02-21-29-16-925.png
>
>
> *Description*
> target file overlap with previous file 
>  
> *Reason*
> In the process of getting the file list and starting to select files, the file list is updated (the file is deleted or the status is updated). When a seq file is occupied and deleted by other compaction thread, and it happens to select files in cross space compaction, the deleted file will be filtered out, resulting in an overlap target file.
>  
> *Solution*
> Do not filter out any seq files before selecting the files. After selecting the files, judge whether the seq files have been deleted or being compacted. If there are any, the task will be discarded.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)