[
https://issues.apache.org/jira/browse/HBASE-13531?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14530786#comment-14530786
]
Jingcheng Du commented on HBASE-13531:
--------------------------------------
Hi Jon, [~jmhsieh].
In the CellSkipListSet, the sequenceId is counted in comparing, so different
cells with the same row key but different seqId can stay in the same
CellSkipListSet.
So I think the violation is caused by the missed readPt when we read cells from
MOB files.
Here is the possible scenario, all the cell has the same row key.
# A.cell0 is added, its seqId is 0.
# A scanner s0 is opend, its readPt is 0, but now start to read this cell0.
# Now another cell1 is added, its seqId is 1.
# Now memstore is flushed, these two cells are in the same mob file.
# The s0 starts to read cell0, and find the mob file.
# Seek the row key in the mob file, we don't use the readPt, so the cell1 is
found.
I think this can happen. Please advise, Jon.
I will look into the code and fix the readPt in mob file.
> After 4/18/15 merge, flakey failures of
> TestAcidGuarantees#testMobScanAtomicity
> -------------------------------------------------------------------------------
>
> Key: HBASE-13531
> URL: https://issues.apache.org/jira/browse/HBASE-13531
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver, Scanners
> Affects Versions: hbase-11339
> Reporter: Jonathan Hsieh
> Fix For: hbase-11339
>
>
> After the merge of master from 4/18/15 with hbase-11339 branch, we encounter
> some atomicity violations. We want to fix before calling merge to trunk.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
