[ https://issues.apache.org/jira/browse/HBASE-18228?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16053314#comment-16053314 ]
Andrew Purtell commented on HBASE-18228: ---------------------------------------- Fine, we can do some of this for 1.x only as appropriate. Another improvement I'd like to suggest is an optional toggle (with a default that can be set in site configuration, so we don't accidentally forget to include it on a command line) for a new preflight check that, after we've disabled balancing, catalog janitor, and splits and merges, ensures - to a reasonable definition of that, given this is a repair tool meant to run in a sick environment - that no splitting or merging activity is in-flight. It's fine if this means that stuck splits cause hbck to stall, hbck can be re-run with a command line flag that forces forward progress if desired. Or the operator could kill the regionserver(s) where the stuck splits are in progress if preferred. > HBCK improvements > ----------------- > > Key: HBASE-18228 > URL: https://issues.apache.org/jira/browse/HBASE-18228 > Project: HBase > Issue Type: Improvement > Components: hbck > Reporter: Lars Hofhansl > Priority: Critical > Fix For: 1.4.0 > > > We just had a prod issue and running HBCK the way we did actually causes more > problems. > In part HBCK did stuff we did not expect, in part we had little visibility > into what HBCK was doing, and in part the logging was confusing. > I'm proposing 2 improvements: > 1. A dry-run mode. Run, and just list what would have been done. > 2. An interactive mode. Run, and for each action request Y/N user input. So > that a user can opt-out of stuff. > [~jmhsieh], FYI -- This message was sent by Atlassian JIRA (v6.4.14#64029)