> On Nov. 20, 2013, 4:16 p.m., kturner wrote: > > test/system/continuous/hdfs-agitator.pl, line 104 > > <https://reviews.apache.org/r/15650/diff/1/?file=388001#file388001line104> > > > > What are the pros and cons of using this haadmin command vs killing > > namenode processes? > > Sean Busbey wrote: > Pro haadmin: > > * The underlying HDFS instance may not be configured for automatic > failover. > * The haadmin command doesn't require knowing where the NameNode > processes are running within the cluster. > * The haadmin tool is a publicly exposed way of saying "do a failover", > whereas finding the NameNode to kill will be a heuristic. > > Pro killing namenode: > > * If you specifically need to test what happens when it's the automatic > failover process kicking in > > Note that I don't think the pro-killing pro is that strong of a pro. The > haadmin command still needs to transition the active to standby and then the > standby to active, so systems above HDFS are going to already encounter e.g. > gaps in there being an active namenode. > > kturner wrote: > I made the following comment on the dev list earlier because review board > was not working. I suspect killing the processes would yield slightly more > realistic test results, but it certainly makes our scripts more unwieldy. > Maybe a better way to do this it to work towards moving hdfs agitation into > hdfs itself. > > Taking things a bit further, killing processes is not as effective in > test as really killing machines (because of it does not expose issues like > unflushed data in OS caches). > > On to another issue. Does the script ever kill all ha namnodes? Is > this possible w/ haadmin?
The kind of testing you're talking about generally happens in BigTop, rather than in individual components (e.g. HBase). Haadmin doesn't have a command to take NameNodes offline, just to mark them as standby rather than active. I believe you could use haadmin to force all namenodes to standby mode, but I'm would suspect in a set up with automatic failover that the failover controllers would cause one to become active again. I'll check to confirm this. Actually getting to the point of killing machines requires something external, e.g. the ability to talk to power managers or VMs. If we're looking for that level of fault testing, then I think we're better off deferring to BigTop and trying to improve both Accumulo's presence there and the use of e.g. Chaos Monkey or Gremlins. - Sean ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/15650/#review29167 ----------------------------------------------------------- On Nov. 18, 2013, 5:13 p.m., Sean Busbey wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/15650/ > ----------------------------------------------------------- > > (Updated Nov. 18, 2013, 5:13 p.m.) > > > Review request for accumulo and Alex Moundalexis. > > > Bugs: ACCUMULO-1794 > https://issues.apache.org/jira/browse/ACCUMULO-1794 > > > Repository: accumulo > > > Description > ------- > > ACCUMULO-1794 adds hdfs failover to continuous integration test. > > > Diffs > ----- > > test/system/continuous/continuous-env.sh.example > 830ae86b5bf2398a840b853423755f6dd65f2dc0 > test/system/continuous/hdfs-agitator.pl PRE-CREATION > test/system/continuous/start-agitator.sh > 52e5a4e82a4564fa624a71f73ad29fa20ba23246 > test/system/continuous/stop-agitator.sh > b853a55b12f8402606af52e0748ca50daf95ed7f > > Diff: https://reviews.apache.org/r/15650/diff/ > > > Testing > ------- > > Ran the hdfs agitator on a CDH4 cluster configured for HA. it successfully > caused the active namenode to failover as it went. > > > Thanks, > > Sean Busbey > >
