[ https://issues.apache.org/jira/browse/HBASE-5360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13253187#comment-13253187 ]
Jonathan Hsieh commented on HBASE-5360: --------------------------------------- @Anoop. Partially. HBASE-5719 sidelines potential overlapping regions but doesn't take .META.'s info about if the region is offline into account. It could sideline a live region while it could have been more efficient to sideline an offline region. However, HBASE-5719 was sufficient for the case we were dealing with. Let's make this a placeholder for ways to improve it. Some ideas include: * Taking a region's offline/splitparent state into account if .META. entry is present * Making decisions about sidelining vs merging based on region size and max region size (instead of range) xml properties. * Improving the heuristic used to decide which region are sidelined. > [uberhbck] Add options for how to handle offline split parents. > ---------------------------------------------------------------- > > Key: HBASE-5360 > URL: https://issues.apache.org/jira/browse/HBASE-5360 > Project: HBase > Issue Type: Improvement > Components: hbck > Affects Versions: 0.90.7, 0.92.1, 0.94.0 > Reporter: Jonathan Hsieh > > In a recent case, we attempted to repair a cluster that suffered from > HBASE-4238 that had about 6-7 generations of "leftover" split data. The hbck > repair options in an development version of HBASE-5128 treat HDFS as ground > truth but didn't check SPLIT and OFFLINE flags only found in meta. The net > effect was that it essentially attempted to merge many regions back into its > eldest geneneration's parent's range. > More safe guards to prevent "mega-merges" are being added on HBASE-5128. > This issue would automate the handling of the "mega-merge" avoiding cases > such as "lingering grandparents". The strategy here would be to add more > checks against .META., and perform part of the catalog janitor's > responsibilities for lingering grandparents. This would potentially include > options to sideline regions, deleting grandparent regions, min size for > sidelining, and mechanisms for cleaning .META.. > Note: There already exists an mechanism to reload these regions -- the bulk > loaded mechanisms in LoadIncrementalHFiles can be used to re-add grandparents > (automatically splitting them if necessary) to HBase. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira