tablet split

[z11373 <z1...@outlook.com>]

As my understanding, Accumulo will have data already sorted with row id, and
if the number of rows is growing, it will split the tablet at one point.
For example, let say I have following row ids:

1_abcxxx
1_abdxxx
1_abexxx
1_abfxxx
1_abgxxx
1_abhxxx
1_abixxx
...
1_zzzxxx
2_abcxxx
2_abdxxx
2_abexxx
2_abfxxx
2_abgxxx
2_abhxxx
...

Let say the data with row id starts with "1_" has a million of rows, and for
sake of example, let say the tablet size is 400K, so in this case the "1_"
data will be split into 3 tablets.

My question is will Accumulo distribute those 3 tablets into different
tablet server nodes? Or perhaps two or all of them will remain in that
original tablet server?


Thanks,
Z




--
View this message in context: http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399.html
Sent from the Developers mailing list archive at Nabble.com.

Re: tablet split

[z11373 <z1...@outlook.com>]

Thanks Josh!
Unfortunately, the monitor only shows which tablet servers where a specific
table's tablets are hosted, and number of entries. I actually need
additional information.
For example, current info shown on the monitor for table A (with 6 tablets):

Tablet Server 1:
# of tablets = 1, # of entries: 10.78M
Tablet Server 2:
# of tablets = 1, # of entries: 22.92M
Tablet Server 3:
# of tablets = 1, # of entries: 13.52M
Tablet Server 4:
# of tablets = 2, # of entries: 135.12M
Tablet Server 5:
# of tablets = 1, # of entries: 34.58M

The data being stored are something like (below is row id value):
1_xxxxxx
.....
.....
2_xxxxxx
.....
.....
3_xxxxxx
.....

There will be case which data with row id 1_xxxxx are ten times more than
2_xxxxx and 3_xxxxx,  which in that case, I am afraid it could be the tablet
server 4 with 135.12M entries (in example above) is where all rows with
'1_xxxxx', or very likely that Accumulo will distribute the data with same
row id prefix to other tablet servers?


Thanks,
Z





--
View this message in context: http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399p15568.html
Sent from the Developers mailing list archive at Nabble.com.

Re: tablet split

[Josh Elser <jo...@gmail.com>]

Quickest check is to just look at the monitor. Click on your table and 
you should be able to drill down into where the tablets are being hosted.

z11373 wrote:
> Back to this thread again...
> If I have 1M of rows with "1_xxxx" data, and 5 tablet servers, how can I
> check if those one million rows got distributed to all tablet servers or
> not? If it doesn't, then I may have perf problem when the client reads the
> data.
>
>
> Thanks,
> Z
>
>
>
> --
> View this message in context: http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399p15536.html
> Sent from the Developers mailing list archive at Nabble.com.

Re: tablet split

[z11373 <z1...@outlook.com>]

Back to this thread again...
If I have 1M of rows with "1_xxxx" data, and 5 tablet servers, how can I
check if those one million rows got distributed to all tablet servers or
not? If it doesn't, then I may have perf problem when the client reads the
data.


Thanks,
Z



--
View this message in context: http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399p15536.html
Sent from the Developers mailing list archive at Nabble.com.

Re: tablet split

[Eric Newton <er...@gmail.com>]

Accumulo will balance the tablets based on the configured balancer.

Without getting down into the details, the splits will be moved to other
nodes.

The Details:

It depends. With the default balancer, it will try to smooth out the number
of tablets among servers, by table. So, if this table goes from 1 tablet to
3, and there are at least 3 servers, each split will eventually find itself
moved to separate server. But, if you add one split among hundreds, it may
not make much of a difference to bother moving the tablet.

-Eric


On Tue, Oct 20, 2015 at 10:39 AM, z11373 <z1...@outlook.com> wrote:

> As my understanding, Accumulo will have data already sorted with row id,
> and
> if the number of rows is growing, it will split the tablet at one point.
> For example, let say I have following row ids:
>
> 1_abcxxx
> 1_abdxxx
> 1_abexxx
> 1_abfxxx
> 1_abgxxx
> 1_abhxxx
> 1_abixxx
> ...
> 1_zzzxxx
> 2_abcxxx
> 2_abdxxx
> 2_abexxx
> 2_abfxxx
> 2_abgxxx
> 2_abhxxx
> ...
>
> Let say the data with row id starts with "1_" has a million of rows, and
> for
> sake of example, let say the tablet size is 400K, so in this case the "1_"
> data will be split into 3 tablets.
>
> My question is will Accumulo distribute those 3 tablets into different
> tablet server nodes? Or perhaps two or all of them will remain in that
> original tablet server?
>
>
> Thanks,
> Z
>
>
>
>
> --
> View this message in context:
> http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399.html
> Sent from the Developers mailing list archive at Nabble.com.
>

Re: tablet split

[Josh Elser <jo...@gmail.com>]

IIRC, either half of a split tablet will remain on the same node as the 
parent; however the next invocation of the configured balancer might 
move them per its policy.

z11373 wrote:
> As my understanding, Accumulo will have data already sorted with row id, and
> if the number of rows is growing, it will split the tablet at one point.
> For example, let say I have following row ids:
>
> 1_abcxxx
> 1_abdxxx
> 1_abexxx
> 1_abfxxx
> 1_abgxxx
> 1_abhxxx
> 1_abixxx
> ...
> 1_zzzxxx
> 2_abcxxx
> 2_abdxxx
> 2_abexxx
> 2_abfxxx
> 2_abgxxx
> 2_abhxxx
> ...
>
> Let say the data with row id starts with "1_" has a million of rows, and for
> sake of example, let say the tablet size is 400K, so in this case the "1_"
> data will be split into 3 tablets.
>
> My question is will Accumulo distribute those 3 tablets into different
> tablet server nodes? Or perhaps two or all of them will remain in that
> original tablet server?
>
>
> Thanks,
> Z
>
>
>
>
> --
> View this message in context: http://apache-accumulo.1065345.n5.nabble.com/tablet-split-tp15399.html
> Sent from the Developers mailing list archive at Nabble.com.