hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Yu <yuzhih...@gmail.com>
Subject Re: Dropping a very large table - 75million rows
Date Thu, 09 Feb 2017 22:28:49 GMT
bq. After the drop, the new table's regions showed a large drop in locality.

Close to 1700 regions were dropped, please check master log for the
duration table was dropped to see how many regions (of the other tables)
were moved.
Region movement might result in drop in locality.

In StochasticLoadBalancer, there is LocalityCostFunction where:

    private static final String LOCALITY_COST_KEY =
"hbase.master.balancer.stochastic.localityCost";

    private static final float DEFAULT_LOCALITY_COST = 25;

The default weight is very low, compared to region count weight (500).

If you want balancer to favor locality when moving regions, please give the
above config higher weight (on same level as 500).


Cheers

On Thu, Feb 9, 2017 at 2:17 PM, Ganesh Viswanathan <gansvv@gmail.com> wrote:

> Thanks Ted/Josh.
>
> Ted-
> I store historical metrics on the locality of regions in each regionserver
> in the cluster. I noticed that the old table had many regions with low
> locality before the drop while the other newer tables had very few cases of
> low locality. After the drop, the new table's regions showed a large drop
> in locality.
>
> Josh-
> I did go from 2200regions before the drop to about 540regions. So,
> relatively speaking, yes it could be that the old table had more regions
> with overall higher locality (though it showed up as more regions in each
> regionserver with lower locality).
>
> At what point (in terms of put/get load and row count/storage size in
> HBase), does it make sense to dive into the data-locality settings and try
> to tune it so that compaction and locality changes are more predictable?
> Would looking at how to create regions (instead of auto-sharding) provide
> greater benefit?
>
> Thanks!
> Ganesh
>
>
> On Thu, Feb 9, 2017 at 11:51 AM, Josh Elser <elserj@apache.org> wrote:
>
> > It could be that the table you dropped had a very good locality while the
> > other tables had less. So, your overall locality went down (when the
> "good"
> > locality regions were no longer included). This wouldn't have affected
> your
> > system's performance because the locality for the table didn't change --
> > just the system-wide locality.
> >
> >
> > Ted Yu wrote:
> >
> >> bq. The locality of regions for OTHER tables on the same regionserver
> also
> >> fell drastically
> >>
> >> Can you be a bit more specific on how you came to the above conclusion ?
> >> Dropping one table shouldn't affect locality of other tables - unless
> >> number of regions on each server becomes unbalanced which triggers
> >> balancer
> >> activities.
> >>
> >> Thanks
> >>
> >> On Thu, Feb 9, 2017 at 7:34 AM, Ganesh Viswanathan<gansvv@gmail.com>
> >> wrote:
> >>
> >> So here is what I observed.
> >>> Dropping this large table had an immediate effect on average locality
> for
> >>> the entire cluster. The locality of regions for OTHER tables on the
> same
> >>> regionserver also fell drastically in the cluster. This was unexpected
> (I
> >>> only thought locality of regions for the dropped table would be
> >>> impacted).
> >>> Is this because of compaction? Does the locality computation use the
> size
> >>> of other regions on each regionserver?
> >>>
> >>> The large drop in locality, however, did not cause latency issues on
> read
> >>> writes for the other tables. Why is that? Is it because I did not try
> to
> >>> hit all low locality regions?
> >>>
> >>> (On another note, I was able to test and perform deletions on per
> region
> >>> basis, but that requires hbck -repair and it seemed more invasive on
> the
> >>> entire cluster health.)
> >>>
> >>> Thanks,
> >>> Ganesh
> >>>
> >>>
> >>> On Sat, Feb 4, 2017 at 11:20 AM Josh Elser<elserj@apache.org>  wrote:
> >>>
> >>> Ganesh,
> >>>>
> >>>> Just drop the table. You are worried about nothing.
> >>>>
> >>>> On Feb 3, 2017 16:51, "Ganesh Viswanathan"<gansvv@gmail.com> 
wrote:
> >>>>
> >>>> Hello Josh-
> >>>>>
> >>>>> I am trying to delete the entire table and recover the disk space.
I
> do
> >>>>>
> >>>> not
> >>>>
> >>>>> need to pick specific contents of the table (if thats what you are
> >>>>>
> >>>> asking
> >>>
> >>>> with #2).
> >>>>> My question is would disabling and dropping such a large table affect
> >>>>>
> >>>> data
> >>>>
> >>>>> locality in a bad way, or slow down the cluster when major_compaction
> >>>>>
> >>>> (or
> >>>
> >>>> whatever cleans up the tombstoned rows) happens. I also read from
> >>>>>
> >>>> another
> >>>
> >>>> post that it can spawn zookeeper transactions and even lock the
> >>>>>
> >>>> zookeeper
> >>>
> >>>> nodes. Is there any concern around zookeeper functionality when
> >>>>>
> >>>> dropping
> >>>
> >>>> large HBase tables.
> >>>>>
> >>>>> Thanks again for taking the time to respond to my questions!
> >>>>>
> >>>>> Ganesh
> >>>>>
> >>>>>
> >>>>>
> >>>>> On Fri, Feb 3, 2017 at 1:12 PM, Josh Elser<elserj@apache.org>
> wrote:
> >>>>>
> >>>>> Ganesh -- I was trying to get at maybe there is a terminology issue
> >>>>>>
> >>>>> here.
> >>>>
> >>>>> If you disable+drop the table, this is an operation on the order
of
> >>>>>>
> >>>>> Regions
> >>>>>
> >>>>>> you have. The number of rows/entries is irrelevant. Closing
and
> >>>>>>
> >>>>> deleting
> >>>>
> >>>>> a
> >>>>>
> >>>>>> region is a relatively fast operation.
> >>>>>>
> >>>>>> Can you please confirm: are you trying to delete the entire
table or
> >>>>>>
> >>>>> are
> >>>>
> >>>>> you trying to delete the *contents* of a table?
> >>>>>>
> >>>>>> If it is the former, I stand by my "you're worried about nothing"
> >>>>>>
> >>>>> comment
> >>>>
> >>>>> :)
> >>>>>>
> >>>>>>
> >>>>>> Ganesh Viswanathan wrote:
> >>>>>>
> >>>>>> Thanks Josh.
> >>>>>>>
> >>>>>>> Also, I realized I didnt give the full size of the table.
It takes
> >>>>>>>
> >>>>>> in
> >>>
> >>>> ~75million rows per minute and stores for 15days. So around
> >>>>>>>
> >>>>>> 1.125billion
> >>>>
> >>>>> rows total.
> >>>>>>>
> >>>>>>> On Fri, Feb 3, 2017 at 12:52 PM, Josh Elser<elserj@apache.org>
> >>>>>>>
> >>>>>> wrote:
> >>>>
> >>>>> I think you are worried about nothing, Ganesh.
> >>>>>>>
> >>>>>>>> If you want to drop (delete) the entire table, just
disable and
> >>>>>>>>
> >>>>>>> drop
> >>>
> >>>> it
> >>>>
> >>>>> from the shell. This operation is not going to have a significant
> >>>>>>>>
> >>>>>>> impact
> >>>>>
> >>>>>> on
> >>>>>>>> your cluster (save a few flush'es). This would only
happen if you
> >>>>>>>>
> >>>>>>> have
> >>>>
> >>>>> had
> >>>>>>>> recent writes to this table (which seems unlikely if
you want to
> >>>>>>>>
> >>>>>>> drop
> >>>
> >>>> it).
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> Ganesh Viswanathan wrote:
> >>>>>>>>
> >>>>>>>> Hello,
> >>>>>>>>
> >>>>>>>>> I need to drop an old HBase table that is quite
large. It has
> >>>>>>>>>
> >>>>>>>> anywhere
> >>>>
> >>>>> between 2million and 70million datapoints. I turned off the count
> >>>>>>>>>
> >>>>>>>> after
> >>>>>
> >>>>>> it
> >>>>>>>>> ran on the HBase shell for half a day. I have 4
other tables that
> >>>>>>>>>
> >>>>>>>> have
> >>>>
> >>>>> around 75million rows in total and also take heavy PUT and GET
> >>>>>>>>>
> >>>>>>>> traffic.
> >>>>>
> >>>>>> What is the best practice for disabling and dropping such a
large
> >>>>>>>>>
> >>>>>>>> table
> >>>>>
> >>>>>> in
> >>>>>>>>> HBase so that I have minimal impact on the rest
of the cluster?
> >>>>>>>>> 1) I hear there are ways to disable (and drop?)
specific regions?
> >>>>>>>>>
> >>>>>>>> Would
> >>>>>
> >>>>>> that work?
> >>>>>>>>> 2) Should I scan and delete a few rows at a time
until the size
> >>>>>>>>>
> >>>>>>>> becomes
> >>>>>
> >>>>>> manageable and then disable/drop the table?
> >>>>>>>>>      If so, what is a good number of rows to delete
at a time,
> >>>>>>>>>
> >>>>>>>> should I
> >>>>
> >>>>> run
> >>>>>>>>> a
> >>>>>>>>> major compaction after these row deletes on specific
regions, and
> >>>>>>>>>
> >>>>>>>> what
> >>>>
> >>>>> is
> >>>>>>>>> a
> >>>>>>>>> good sized table that can be easily dropped (and
has been
> >>>>>>>>>
> >>>>>>>> validated)
> >>>
> >>>> without causing issues on the larger cluster.
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> Thanks!
> >>>>>>>>> Ganesh
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>
> >>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message