[
https://issues.apache.org/jira/browse/HBASE-20933?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vishal Khandelwal updated HBASE-20933:
--------------------------------------
Attachment: Test123.java
> multiple splits may result into forever uncleaned split region
> --------------------------------------------------------------
>
> Key: HBASE-20933
> URL: https://issues.apache.org/jira/browse/HBASE-20933
> Project: HBase
> Issue Type: Bug
> Reporter: Vishal Khandelwal
> Assignee: Vishal Khandelwal
> Priority: Major
> Attachments: Test123.java
>
>
> Incase of multiple subsequent split and with an open handle on old reference
> file, it may result into split region which can never be cleaned
> So Here are two issues.
> # Region is getting split even when it has reference to its parent
> # Region is going offline/in archive mode even when there are reference
> pending in store
> *Repro Steps*
> # Region split (P)
> # Before major compaction starts after split, open a handle on store file on
> new region (DA & DB)
> # Let compaction completes on DA, (Here compaction will not clear reference
> store files as it is opened)
> # Split new region (DA) again ( shouldSplit will return true as before
> compaction even does the cleanup, it removes the compacted files and
> reference in-memory list)
> # Now CatalogJanitor will not remove this region as it has store references,
> majorCompaction/CompactedHFilesDischarger will not do the cleanup as it only
> looks at only online regions
> # After above steps region-DA which is offline will always be in split
> regions and never getting cleaned up.
> We found that catalog janitor is also not able to clean regions which are
> offline(split parent) because it has reference of the daughter of it's parent
> which is not getting cleaned up. This is causing lot of store files not
> getting cleaned causing more space in local index store and lot of split
> lingering regions.
> Unit test repro the scenario has been attached.
> Fix can be in CompactedHFilesDischarger to look into offline region or stop
> split in such cases.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
