huaxiang sun commented on HBASE-16578:

I did a quick test locally and think it is valid. Please confirm this case.

I reviewed  Jingcheng's fix. I have one case which I think it will not work.

1. put mob cell r1, flush, it will create ref1 and mobFile1.
2. put mob cell r2, flush, it will create ref2 and mobFile2.
3. put normal cell r3, flush, it will create hfile3.
4. minor or major compact, ref1, ref2, hfile3, it creates hfile4.
5. mob compact, it will create mobfile3 and bulkload ref3.
6. compact hfile4 and ref3, then again ref cells in hfile4 will win since 
hfile4 got a bigger maxSeqId.



> Mob data loss after mob compaction and normal compcation
> --------------------------------------------------------
>                 Key: HBASE-16578
>                 URL: https://issues.apache.org/jira/browse/HBASE-16578
>             Project: HBase
>          Issue Type: Bug
>          Components: mob
>    Affects Versions: 2.0.0
>            Reporter: huaxiang sun
>            Assignee: Jingcheng Du
>         Attachments: HBASE-16578-V2.patch, HBASE-16578.patch, 
> TestMobCompaction.java, TestMobCompaction.java
> StoreFileScanners on MOB cells rely on the scannerOrder to find the latest 
> cells after mob compaction. The value of scannerOrder is assigned by the 
> order of maxSeqId of StoreFile, and this maxSeqId is valued only after the 
> reader of the StoreFile is created.
> In {{Compactor.compact}}, the compacted store files are cloned and their 
> readers are not created. And in {{StoreFileScanner.getScannersForStoreFiles}} 
> the StoreFiles are sorted before the readers are created and at that time the 
> maxSeqId for each file is -1 (the default value). This will lead  to a chaos 
> in scanners in the following normal compaction. Some older cells might be 
> chosen during the normal compaction.
> We need to create readers either before the sorting in the method 
> {{StoreFileScanner.getScannersForStoreFiles}}, or create readers just after 
> the store files are cloned in {{Compactor.compact}}.

This message was sent by Atlassian JIRA

Reply via email to