[ https://issues.apache.org/jira/browse/HBASE-21355?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16658527#comment-16658527 ]
Duo Zhang commented on HBASE-21355: ----------------------------------- Talked with [~openinx] offline, we think that the problem is the code introduced in HBASE-20940. We should not try to open the store files every time when calling getRegionInfo, and then close them. It is too expensive. Instead, I think we could also check the compacted files when testing whether a region is mergable or splittable. And I think we should provide a UT for this, as there is no test for HBASE-20940. > HStore's storeSize is calculated repeatedly which causing the confusing > region split > ------------------------------------------------------------------------------------- > > Key: HBASE-21355 > URL: https://issues.apache.org/jira/browse/HBASE-21355 > Project: HBase > Issue Type: Bug > Components: regionserver > Reporter: Zheng Hu > Assignee: Zheng Hu > Priority: Blocker > Fix For: 3.0.0, 1.5.0, 1.3.3, 2.2.0, 2.1.1, 2.0.3, 1.4.9, 1.2.9 > > Attachments: HBASE-21355.branch-1.patch, HBASE-21355.v1.patch > > > When testing the branch-2's write performance in our internal cluster, we > found that the region will be inexplicably split. > We use the default ConstantSizeRegionSplitPolicy and > hbase.hregion.max.filesize=40G,but the region will be split even if its > bytes size is less than 40G(only ~6G). > Checked the code, I found that the following path will accumulate the > store's storeSize to a very big value, because the path has no reset.. > {code} > RsRpcServices#getRegionInfo > -> HRegion#isMergeable > -> HRegion#hasReferences > -> HStore#hasReferences > -> HStore#openStoreFiles > {code} > BTW, we seems forget to maintain the read replica's storeSize when refresh > the store files. -- This message was sent by Atlassian JIRA (v7.6.3#76005)