Mailing-List: contact user-help@hbase.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@hbase.apache.org
Received-SPF: pass (nike.apache.org: local policy)
DomainKey-Signature: a=rsa-sha1; q=dns; c=nofws;
  s=s1024; d=yahoo.com;
  h=Message-ID:X-YMail-OSG:Received:X-Mailer:References:Date:From:Subject:To:In-Reply-To:MIME-Version:Content-Type;
  b=FNXcmQMx8RtoZAXS+XrqzzI2IGebgQ+Cfj52Qjxg4h9sz4eOVl0nw0OUVf9Iv+PRVbeOPar3yzeiZOc2maTiTWbeFGvbeaKuzj9wuWikMcr/Wxgv52GInsqnayENPYXF1C3zdbku/Pc7BBUQSQlnNCGh4igqWOg5Uqh/BT2rVw8=;
Message-ID: <137933.34958.qm@web137320.mail.in.yahoo.com>
References: <692424.89426.qm@web137319.mail.in.yahoo.com>
 <AANLkTi=S+bxaCQ_yHc9MtLqdWdPU+CRKQcZRc_zRATpA@mail.gmail.com>
 <747855.89176.qm@web137312.mail.in.yahoo.com>
 <AANLkTi=6Gu0OiKg__89uv9nLa3FiKnZpjLU2sVQ62XKW@mail.gmail.com>
 <856663.46924.qm@web137315.mail.in.yahoo.com>
 <AANLkTi=TWctwu7e96NAA95iL8qTnOVaZKV29t1cKg81n@mail.gmail.com>
Date: Mon, 6 Sep 2010 10:32:05 +0530 (IST)
From: "Murali Krishna. P" <muralikpbhat@yahoo.com>
Subject: Re: HBase secondary index performance
To: user@hbase.apache.org
In-Reply-To: <AANLkTi=TWctwu7e96NAA95iL8qTnOVaZKV29t1cKg81n@mail.gmail.com>
MIME-Version: 1.0
Content-Type: multipart/alternative; boundary="0-1947991794-1283749325=:34958"

--0-1947991794-1283749325=:34958
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable

Hi,=0A   My row size is around 300 bytes with total 20 columns. I tried the=
 custom =0Aindexing without the write to WAL. Currently having only 2 table=
s, one for the =0Amain table and another for all 20 indexes. My key to the =
index table is =0AcolumnValue+columnName+rowKey.=0AI am getting around 500 =
inserts/second now. (ie, total of ~10K puts). This is =0Aprobably comparabl=
e with your numbers based on the data size.=0A  I have some doubts with the=
 hbase write implementation. =0A* Is this the max that we can achieve with =
any number of region servers? Why =0Aadding region servers not improving th=
e write performance? Is it because when =0Athe data doesn't exist in the ta=
ble, it always writes to one region ? =0A=0A* Probably writing to an existi=
ng, well distributed table might give better =0Aperformance since the write=
s will be across machines ? In that case, if we have =0Amultiple tables (on=
e per index), will it be better during the initial write =0Aitself (since e=
ach table has different region) ??=0A=0A Thanks,=0AMurali Krishna=0A=0A=0A=
=0A=0A________________________________=0AFrom: Andrey Stepachev <octo47@gma=
il.com>=0ATo: user@hbase.apache.org=0ASent: Sun, 5 September, 2010 11:54:45=
 PM=0ASubject: Re: HBase secondary index performance=0A=0A2010/9/5 Murali K=
rishna. P <muralikpbhat@yahoo.com>:=0A> Hi,=0A>        Thanks for the detai=
led explanation, I liked the idea of timestamp=0A> check, this will be good=
 enough for us and I can put a periodic MR cleaner.=0A> However I need some=
 help in understanding the 30K number that was claimed.=0A=0AReal insert ra=
te will depend on size of row, size of write buffer etc.=0AIn case of simpl=
e row with one long  per row i got 30k requests/second=0A(shown in hbase).=
=0AJson serialised objects 100-700bytes each, with validation I can insert =
2-6k=0Aobjects (json) per second.=0A=0AWith=0A> the IndexedTable approach, =
I got only 1200rows/s (60rows/s X 20 index =0Acolumns).=0A> I understood th=
at there arean additional reads that indextable does but  25X=0A> improveme=
nt that you got is very impressive. Can you please help me to=0A> understan=
d this gain ? (My hardware is 8GB/7.2rpm/2core-2GHz)=0A=0ADid you try to in=
sert data into non indexed region (disable=0Aindexedtables extension)?=0AWh=
at numbers did you got?=0A=0A>=0A>  Thanks,=0A> Murali Krishna=0A>=0A>=0A>=
=0A>=0A> ________________________________=0A> From: Andrey Stepachev <octo4=
7@gmail.com>=0A> To: user@hbase.apache.org=0A> Sent: Sun, 5 September, 2010=
 3:53:26 AM=0A> Subject: Re: HBase secondary index performance=0A>=0A> 2010=
/9/3 Murali Krishna. P <muralikpbhat@yahoo.com>:=0A>=0A>>        * custom i=
ndexing is good, but our data keeps changing every day. So,=0A>>probably=0A=
>> indextable is the best option for us=0A>=0A> In case of custom indexing =
you can use timestamps to check, that index=0A> record still valid.=0A> (or=
 ever simply recheck existance of the value)=0A> Also you need regular inde=
x cleanup (mr job or some custom application).=0A>=0A> To index some row id=
entified by 'key' having 'value', we can create=0A> index table,=0A> where =
key will be [value:key] and insert rows every time, when we insert=0A> our =
values. We will got 30k rows/s/node.=0A> When we want to find all 'value', =
we scan [value:0000, value:9999] and=0A> find all keys,=0A> which point to =
rows, containing values.=0A> We scan index, random get rows, recheck, that =
index is still valid=0A> (check value or timestamp, index timestamp should =
be >=3D value timestamp) and=0A> return only valid values (may be we can ev=
en delete on the fly when we=0A> got negative=0A> result to automatically c=
lenup stale data).=0A>=0A>=0A>>        * Just added one more regionserver a=
nd it did not help. Actually it =0A>went=0A>>back=0A>> to 60/s for some str=
ange reason(with one client). The requests in the hbase =0Aui=0A>> is not u=
niform across 2 region servers. One server is doing around 2000 and=0A> the=
=0A>> other 500. Probably once the region gets split and when we have lots =
of data,=0A>> writes will improve ? (Now it is just writing to one region f=
or the main=0A> table)=0A>=0A> Looks like all data goes to one region serve=
r. Try to make more random writes=0A> (may be you should make key as random=
 uuid or other key randomization =0A>technique)=0A>=0A>>        * Is there =
some way to do bulk load the indexedtable? Earlier I have=0A>>used the=0A>>=
 bulk loader tool (mapreduce job which creates the regions offline) but not=
=0A> sure=0A>> whether it works with indexed table.=0A>=0A> No sure, but yo=
u can look at source code, and try to emulate indexing=0A> operations in=0A=
> your code after regular bulk loading.=0A>=0A>>=0A>>=0A>>  Thanks,=0A>> Mu=
rali Krishna=0A>>=0A>>=0A>=0A> Andrey.=0A>=0A
--0-1947991794-1283749325=:34958--