considering 80-20 rule(does it apply to hbase tables?), what about specifying balance property(per-table vs non per-table) on a table basis(default is non per-table), then those few large tables could be balanced well with acceptable total balance cost?
On Tue, Sep 23, 2014 at 11:35 AM, Qiang Tian <tian...@gmail.com> wrote: > Hello, I happened to got balancer related issues 2 months ago and looked > at that part, below is a summary: > 1)by default, hbase balancer(StochasticLoadBalancer by default) does not > balance regions per table. i.e. all regions are considered as 1 table. so > if you have many tables, especially some tables have empty regions, you > probably get unbalanced, the balancer probably not triggered at all. > this is got from code inspection, my problem failed to be reproduced > later. but it proved that deleting empty regions can trigger balancer > correctly and make regions well balanced. > > 2)there are some other reasons that balancer are not triggered. see > HMaster#balance. turn on debug can see related messages in master log. in > my case, it is not triggered because there are regions in transition: > LOG.debug("Not running balancer because " + regionsInTransition.size() + > " region(s) in transition: " + > org.apache.commons.lang.StringUtils. > abbreviate(regionsInTransition.toString(), 256)); > > the cause can be found in regionserver log file. > > 3)per-table balance can be set by "hbase.master.loadbalance.bytable", > however it looks not a good option when you have many tables - the master > will issue balance call for each table, one by one. > > 4)split region follows normal balancer process. so if you have issue in > #1, split does not help balance. it looks pre-split at table creation is > fine, which uses round-robin assignment. > > > > On Tue, Sep 23, 2014 at 2:12 AM, Bharath Vissapragada < > bhara...@cloudera.com> wrote: > >> https://issues.apache.org/jira/browse/HBASE-11368 related to the original >> issue too. >> >> On Mon, Sep 22, 2014 at 10:18 AM, Ted Yu <yuzhih...@gmail.com> wrote: >> >> > As you noted in the FIXME, there're some factors which should be >> tackled by >> > balancer / assignment manager. >> > >> > Please continue digging up master log so that we can find the cause for >> > balancer not fulfilling your goal. >> > >> > Cheers >> > >> > On Mon, Sep 22, 2014 at 10:09 AM, Jianshi Huang < >> jianshi.hu...@gmail.com> >> > wrote: >> > >> > > Ok, I fixed this by manually reassign region servers to newly created >> > ones. >> > > >> > > def reassignRegionServer(admin: HBaseAdmin, regions: >> Seq[HRegionInfo], >> > > regionServers: Seq[ServerName]): Unit = { >> > > val rand = new Random() >> > > regions.foreach { r => >> > > val idx = rand.nextInt(regionServers.size) >> > > val server = regionServers(idx) >> > > // FIXME: what if selected region server is dead? >> > > admin.move(r.getEncodedNameAsBytes, >> > > server.getServerName.getBytes("UTF8")) >> > > } >> > > } >> > > >> > > er... >> > > >> > > Jianshi >> > > >> > > On Tue, Sep 23, 2014 at 12:24 AM, Jianshi Huang < >> jianshi.hu...@gmail.com >> > > >> > > wrote: >> > > >> > > > Hmm...any workaround? I only want to do this: >> > > > >> > > > Rebalance the new regions *evenly* to all servers after manually >> adding >> > > > splits, so later bulk insertions won't cause contention. >> > > > >> > > > P.S. >> > > > Looks like two of the region servers which had majority of the >> regions >> > > > were down during Major compaction... I guess it had too much data. >> > > > >> > > > >> > > > Jianshi >> > > > >> > > > On Tue, Sep 23, 2014 at 12:13 AM, Jianshi Huang < >> > jianshi.hu...@gmail.com >> > > > >> > > > wrote: >> > > > >> > > >> Yes, I have access to Master UI, however logs/*.log cannot be >> opened >> > or >> > > >> downloaded, must be some security restrictions in the proxy... >> > > >> >> > > >> Jianshi >> > > >> >> > > >> On Tue, Sep 23, 2014 at 12:06 AM, Ted Yu <yuzhih...@gmail.com> >> wrote: >> > > >> >> > > >>> Do you have access to Master UI ? >> > > >>> >> > > >>> <master-address>:60010/logs/ would show you list of log files. >> > > >>> >> > > >>> The you can view >> > > <master-address>:60010/logs/hbase-<user>-master-XXX.log >> > > >>> >> > > >>> Cheers >> > > >>> >> > > >>> On Mon, Sep 22, 2014 at 9:00 AM, Jianshi Huang < >> > > jianshi.hu...@gmail.com> >> > > >>> wrote: >> > > >>> >> > > >>> > Ah... I don't have access to HMaster logs... I need to ask the >> > admin. >> > > >>> > >> > > >>> > Jianshi >> > > >>> > >> > > >>> > On Mon, Sep 22, 2014 at 11:49 PM, Ted Yu <yuzhih...@gmail.com> >> > > wrote: >> > > >>> > >> > > >>> > > bq. assign per-table balancer class >> > > >>> > > >> > > >>> > > No that I know of. >> > > >>> > > Can you pastebin master log involving output from balancer ? >> > > >>> > > >> > > >>> > > Cheers >> > > >>> > > >> > > >>> > > On Mon, Sep 22, 2014 at 8:29 AM, Jianshi Huang < >> > > >>> jianshi.hu...@gmail.com> >> > > >>> > > wrote: >> > > >>> > > >> > > >>> > > > Hi Ted, >> > > >>> > > > >> > > >>> > > > I moved setBalancerRunning before balancer and run them >> twice. >> > > >>> However >> > > >>> > I >> > > >>> > > > still got highly skewed region distribution. >> > > >>> > > > >> > > >>> > > > I guess it's because of the StochasticLoadBalancer, can I >> > assign >> > > >>> > > per-table >> > > >>> > > > balancer class in HBase? >> > > >>> > > > >> > > >>> > > > >> > > >>> > > > Jianshi >> > > >>> > > > >> > > >>> > > > On Mon, Sep 22, 2014 at 9:50 PM, Ted Yu < >> yuzhih...@gmail.com> >> > > >>> wrote: >> > > >>> > > > >> > > >>> > > > > admin.setBalancerRunning() call should precede the call to >> > > >>> > > > > admin.balancer(). >> > > >>> > > > > >> > > >>> > > > > You can inspect master log to see whether regions are >> being >> > > >>> moved off >> > > >>> > > the >> > > >>> > > > > heavily loaded server. >> > > >>> > > > > >> > > >>> > > > > Cheers >> > > >>> > > > > >> > > >>> > > > > On Mon, Sep 22, 2014 at 1:42 AM, Jianshi Huang < >> > > >>> > > jianshi.hu...@gmail.com> >> > > >>> > > > > wrote: >> > > >>> > > > > >> > > >>> > > > > > Hi Ted and others, >> > > >>> > > > > > >> > > >>> > > > > > I did the following after adding splits (without data) >> to >> > my >> > > >>> table, >> > > >>> > > > > however >> > > >>> > > > > > the region is still very imbalanced (one region server >> has >> > > 221 >> > > >>> > > regions >> > > >>> > > > > and >> > > >>> > > > > > other 50 region servers have about 4~8 regions each). >> > > >>> > > > > > >> > > >>> > > > > > admin.balancer() >> > > >>> > > > > > admin.setBalancerRunning(true, true) >> > > >>> > > > > > >> > > >>> > > > > > The balancer class in my HBase cluster is >> > > >>> > > > > > >> > > >>> > > > > > >> > > org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer >> > > >>> > > > > > >> > > >>> > > > > > So, is this behavior expected? Can I assign different >> > > balancer >> > > >>> > class >> > > >>> > > to >> > > >>> > > > > my >> > > >>> > > > > > tables (I don't have HBase admin permission)? Which one >> > > should >> > > >>> I >> > > >>> > use? >> > > >>> > > > > > >> > > >>> > > > > > I just want HBase to evenly distribute the regions even >> > they >> > > >>> don't >> > > >>> > > have >> > > >>> > > > > > data (that's the purpose of pre-split I think). >> > > >>> > > > > > >> > > >>> > > > > > >> > > >>> > > > > > Jianshi >> > > >>> > > > > > >> > > >>> > > > > > >> > > >>> > > > > > On Sat, Sep 6, 2014 at 12:45 AM, Ted Yu < >> > yuzhih...@gmail.com >> > > > >> > > >>> > wrote: >> > > >>> > > > > > >> > > >>> > > > > > > Yes. See the following method in HBaseAdmin: >> > > >>> > > > > > > >> > > >>> > > > > > > public boolean balancer() >> > > >>> > > > > > > >> > > >>> > > > > > > >> > > >>> > > > > > > On Fri, Sep 5, 2014 at 9:38 AM, Jianshi Huang < >> > > >>> > > > jianshi.hu...@gmail.com >> > > >>> > > > > > >> > > >>> > > > > > > wrote: >> > > >>> > > > > > > >> > > >>> > > > > > > > Thanks Ted! >> > > >>> > > > > > > > >> > > >>> > > > > > > > Didn't know I still need to run the 'balancer' >> command. >> > > >>> > > > > > > > >> > > >>> > > > > > > > Is there a way to do it programmatically? >> > > >>> > > > > > > > >> > > >>> > > > > > > > Jianshi >> > > >>> > > > > > > > >> > > >>> > > > > > > > >> > > >>> > > > > > > > >> > > >>> > > > > > > > On Sat, Sep 6, 2014 at 12:29 AM, Ted Yu < >> > > >>> yuzhih...@gmail.com> >> > > >>> > > > wrote: >> > > >>> > > > > > > > >> > > >>> > > > > > > > > After splitting the region, you may need to run >> > > balancer >> > > >>> to >> > > >>> > > > spread >> > > >>> > > > > > the >> > > >>> > > > > > > > new >> > > >>> > > > > > > > > regions out. >> > > >>> > > > > > > > > >> > > >>> > > > > > > > > Cheers >> > > >>> > > > > > > > > >> > > >>> > > > > > > > > >> > > >>> > > > > > > > > On Fri, Sep 5, 2014 at 9:25 AM, Jianshi Huang < >> > > >>> > > > > > jianshi.hu...@gmail.com >> > > >>> > > > > > > > >> > > >>> > > > > > > > > wrote: >> > > >>> > > > > > > > > >> > > >>> > > > > > > > > > Hi Shahab, >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > I see, that seems to be the right way... >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > On Sat, Sep 6, 2014 at 12:21 AM, Shahab Yunus < >> > > >>> > > > > > > shahab.yu...@gmail.com> >> > > >>> > > > > > > > > > wrote: >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > > Shahab >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > -- >> > > >>> > > > > > > > > > Jianshi Huang >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > > LinkedIn: jianshi >> > > >>> > > > > > > > > > Twitter: @jshuang >> > > >>> > > > > > > > > > Github & Blog: http://huangjs.github.com/ >> > > >>> > > > > > > > > > >> > > >>> > > > > > > > > >> > > >>> > > > > > > > >> > > >>> > > > > > > > >> > > >>> > > > > > > > >> > > >>> > > > > > > > -- >> > > >>> > > > > > > > Jianshi Huang >> > > >>> > > > > > > > >> > > >>> > > > > > > > LinkedIn: jianshi >> > > >>> > > > > > > > Twitter: @jshuang >> > > >>> > > > > > > > Github & Blog: http://huangjs.github.com/ >> > > >>> > > > > > > > >> > > >>> > > > > > > >> > > >>> > > > > > >> > > >>> > > > > > >> > > >>> > > > > > >> > > >>> > > > > > -- >> > > >>> > > > > > Jianshi Huang >> > > >>> > > > > > >> > > >>> > > > > > LinkedIn: jianshi >> > > >>> > > > > > Twitter: @jshuang >> > > >>> > > > > > Github & Blog: http://huangjs.github.com/ >> > > >>> > > > > > >> > > >>> > > > > >> > > >>> > > > >> > > >>> > > > >> > > >>> > > > >> > > >>> > > > -- >> > > >>> > > > Jianshi Huang >> > > >>> > > > >> > > >>> > > > LinkedIn: jianshi >> > > >>> > > > Twitter: @jshuang >> > > >>> > > > Github & Blog: http://huangjs.github.com/ >> > > >>> > > > >> > > >>> > > >> > > >>> > >> > > >>> > >> > > >>> > >> > > >>> > -- >> > > >>> > Jianshi Huang >> > > >>> > >> > > >>> > LinkedIn: jianshi >> > > >>> > Twitter: @jshuang >> > > >>> > Github & Blog: http://huangjs.github.com/ >> > > >>> > >> > > >>> >> > > >> >> > > >> >> > > >> >> > > >> -- >> > > >> Jianshi Huang >> > > >> >> > > >> LinkedIn: jianshi >> > > >> Twitter: @jshuang >> > > >> Github & Blog: http://huangjs.github.com/ >> > > >> >> > > > >> > > > >> > > > >> > > > -- >> > > > Jianshi Huang >> > > > >> > > > LinkedIn: jianshi >> > > > Twitter: @jshuang >> > > > Github & Blog: http://huangjs.github.com/ >> > > > >> > > >> > > >> > > >> > > -- >> > > Jianshi Huang >> > > >> > > LinkedIn: jianshi >> > > Twitter: @jshuang >> > > Github & Blog: http://huangjs.github.com/ >> > > >> > >> >> >> >> -- >> Bharath Vissapragada >> <http://www.cloudera.com> >> > >