[ https://issues.apache.org/jira/browse/HBASE-12629?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14235192#comment-14235192 ]
Jonathan Hsieh commented on HBASE-12629: ---------------------------------------- For users or operators, if a feature isn't documented it doesn't exist and the default is the only behavior. That's one vote for removing the knob. I went and rereviewed HBASE-10413 -- the patch that introduces the RegionSizeCalculator and adds the info the the InputSplit's getLength() call. This doesn't affect the NN -- the data is gathered from ClusterStats -- at most an rpc that gathers metric info from the HMaster. (I guess I can see a case in the futre when this could be a problem). As a nit, implementation-wise the check on whether to do anything useful is inside the RegionSizeCalculator seems wrong. Instead of having RegionSizeCalculator decide internally whether it is enabled/disabled, it should be done by either: 1) removing the knob so that you always have the ClusterStatus based RegionSizeCalculator (the first patch here) 2) having an "AlwaysZeroRegionSizeCalculator" and a "ClusterStatusRegionSizeCaculator" and leave making the decision logic for the TableInputFormatBase/MultiTableInputFormat code. FYI, the patch I'm looking at -- HBASE-12590 -- rejiggers MR InputSplit ranges to deskew region splits for MR jobs and explicitly relies on the region size calculation. Having the ClusterStatus based RegionSizeCalc on is a prereq and leaving the ability disable it adds a odd corner case and more testing associated with it. Seems better to just have one knob for this. This leads to: 1) Just removing that rscalc config so we have one knob using the the HBASE-12590 deskew is used. 2) An alternative where we refactor instead of remove and have the HBASE-12590 patch decides on the to pick the "ClusterStatusRegionSizeCalculator" if enabled or "AlwaysZeroRSCalc" if not enabled. We'd combine HBASE-12590 and HBASE-10413's features into one config knob (instead of what currently would be two). If after some testing HBASE-12590 works nicely, I think I'd want it on by default as well. WDYT? > Remove hbase.regionsizecalculator.enable from RegionSizeCalculator > ------------------------------------------------------------------ > > Key: HBASE-12629 > URL: https://issues.apache.org/jira/browse/HBASE-12629 > Project: HBase > Issue Type: Improvement > Reporter: Jonathan Hsieh > Assignee: Jonathan Hsieh > Fix For: 1.0.0, 2.0.0 > > Attachments: > 0001-HBASE-12629-Remove-hbase.regionsizecalculator.enable.patch > > > The RegionSizeCalculator has a option to disable it. It is on by default and > end-to-end use with it disabled is not tested or used anywhere except for a > simple unit test. This removes it. -- This message was sent by Atlassian JIRA (v6.3.4#6332)