[DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Dongjoon Hyun <do...@apache.org>]

Hi, all.

I have a question in these day.
*Can we improve Tajo performance further than now?*
Since Tajo is well architectured, each module can be improved
independently.
>From the SQL parser to the network part, there will be many candidates to
improve.

If you didn't try due to lack of time to implement, you can share your idea.
As a Tajo community member, please feel free to comment anything.
We can discuss and create more issues for that.

When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up
to now 26 queries) and Tajo speed becomes more faster. It depends on our
coding speed. :-)

Best regards,
Dongjoon.

PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
Apache Derby and HBase, http://www.splicemachine.com/ )

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Dongjoon Hyun <do...@apache.org>]

Ur, I mean Tajo is focusing on features, not for a specific benchmark.
I have no other intention such a thing you mentioned.
If it looks like that, it's my fault due to English writing skill.

As you said in this email, WITH/IN/EXIST/scalar subquery will almost finish
TPC-DS.
I knew that. Moreover, you advised me to make TPC-DS wiki page, too.
So, I 'GUESSED' and 'BELIEVE' in the first email in this thread.

I know you and Tajo community consider TPC-DS for a long time before.
(according to the Jira issues.)
What I wanted here is to know Tajo maturity clearly and officially.
I thought Tajo can do that proudly.

Warmly,
Dongjoon.

On Sun, Mar 22, 2015 at 1:51 PM, Hyunsik Choi <hy...@apache.org> wrote:

> TPC-DS is just a benchmark, and it is not test sets. For just test
> cases, why should we aim at TPC-DS? They are not proper as well as are
> overkill for this purpose.
>
> > For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap already.
>
> > TPC-DS does not mean anything to TAJO roadmap.
>
> They are NOT TRUE. As you already know, WITH, IN/EXIST subquery, and
> scalar subquery are already on our roadmap. If they are supported,
> most of TPC-DS queries will work naturally. I'm concerning with your
> intension, saying we do not consider TPC-DS.
>
> On Sat, Mar 21, 2015 at 4:38 PM, Dongjoon Hyun <do...@apache.org>
> wrote:
> > Hyunsik, TPC-DS is not important. It's just a way to see Tajo Maturity
> as a
> > SQL engine. TPC-DS does not mean anything to TAJO roadmap.
> >
> > More specifically, the importance is the same with TPC-H, however Tajo
> > handled TPC-H well now, doesn't it? So, we need just another testsuite to
> > see how Tajo handle SQL(including WITH/IN/EXISTS) clause well. I think
> you
> > agree that in TAJO-1410.
> >
> > TPC defines
> > - TPC-H: TPC-H is an ad-hoc, decision support benchmark.
> > - TPC-DS: TPC Benchmark™DS (TPC-DS): The New Decision Support Benchmark
> > Standard
> >
> > In these days, TPC-DS is replacing the ancient TPC-H from real site
> > customers.
> > The main reason is based on the lack of TPC-H. TPC-H became a basic
> level.
> >
> > Could you tell me if you have some other testcases in mind, except
> TPC-DS?
> >
> > Dongjoon.
> >
> > On Sun, Mar 22, 2015 at 3:58 AM, Hyunsik Choi <hy...@apache.org>
> wrote:
> >
> >> Could you tell me why you think TPC-DS is important?
> >>
> >> On Saturday, March 21, 2015, Dongjoon Hyun <do...@apache.org> wrote:
> >>
> >> > Oh, that was an expression of my regret about my weak pull-requests.
> >> There
> >> > exists no offence. It's a duty. I hope you didn't waste your time much
> >> due
> >> > to my stupid learning cost.
> >> >
> >> > After Hyunsik's direct advice, I have studied Tajo Jira more and
> >> understand
> >> > what you mean slightly more. Now I understand many speedup issues,
> e.g.
> >> > offheap-based scanner, zero-copy physical layer processing. For
> newbies
> >> > like me, we need more men like you.
> >> >
> >> > For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap
> >> already. I
> >> > respect the current roadmap. You need not worry about that. It's my
> >> > personal hope as I mentioned before in this mailing list. Here is more
> >> > recent another report about TPC-DS. (It's consistent with the report
> >> shared
> >> > by Hyunsik.)
> >> >
> >> >
> >> >
> >>
> https://developer.ibm.com/hadoop/blog/2014/12/02/big-sql-3-0-hadoop-ds-benchmark-performance-isnt-everything/
> >> >
> >> >
> >> > Thank you all.
> >> >
> >> > Warmly,
> >> > Dongjoon.
> >> >
> >>
>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

TPC-DS is just a benchmark, and it is not test sets. For just test
cases, why should we aim at TPC-DS? They are not proper as well as are
overkill for this purpose.

> For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap already.

> TPC-DS does not mean anything to TAJO roadmap.

They are NOT TRUE. As you already know, WITH, IN/EXIST subquery, and
scalar subquery are already on our roadmap. If they are supported,
most of TPC-DS queries will work naturally. I'm concerning with your
intension, saying we do not consider TPC-DS.

On Sat, Mar 21, 2015 at 4:38 PM, Dongjoon Hyun <do...@apache.org> wrote:
> Hyunsik, TPC-DS is not important. It's just a way to see Tajo Maturity as a
> SQL engine. TPC-DS does not mean anything to TAJO roadmap.
>
> More specifically, the importance is the same with TPC-H, however Tajo
> handled TPC-H well now, doesn't it? So, we need just another testsuite to
> see how Tajo handle SQL(including WITH/IN/EXISTS) clause well. I think you
> agree that in TAJO-1410.
>
> TPC defines
> - TPC-H: TPC-H is an ad-hoc, decision support benchmark.
> - TPC-DS: TPC Benchmark™DS (TPC-DS): The New Decision Support Benchmark
> Standard
>
> In these days, TPC-DS is replacing the ancient TPC-H from real site
> customers.
> The main reason is based on the lack of TPC-H. TPC-H became a basic level.
>
> Could you tell me if you have some other testcases in mind, except TPC-DS?
>
> Dongjoon.
>
> On Sun, Mar 22, 2015 at 3:58 AM, Hyunsik Choi <hy...@apache.org> wrote:
>
>> Could you tell me why you think TPC-DS is important?
>>
>> On Saturday, March 21, 2015, Dongjoon Hyun <do...@apache.org> wrote:
>>
>> > Oh, that was an expression of my regret about my weak pull-requests.
>> There
>> > exists no offence. It's a duty. I hope you didn't waste your time much
>> due
>> > to my stupid learning cost.
>> >
>> > After Hyunsik's direct advice, I have studied Tajo Jira more and
>> understand
>> > what you mean slightly more. Now I understand many speedup issues, e.g.
>> > offheap-based scanner, zero-copy physical layer processing. For newbies
>> > like me, we need more men like you.
>> >
>> > For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap
>> already. I
>> > respect the current roadmap. You need not worry about that. It's my
>> > personal hope as I mentioned before in this mailing list. Here is more
>> > recent another report about TPC-DS. (It's consistent with the report
>> shared
>> > by Hyunsik.)
>> >
>> >
>> >
>> https://developer.ibm.com/hadoop/blog/2014/12/02/big-sql-3-0-hadoop-ds-benchmark-performance-isnt-everything/
>> >
>> >
>> > Thank you all.
>> >
>> > Warmly,
>> > Dongjoon.
>> >
>>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Dongjoon Hyun <do...@apache.org>]

Hyunsik, TPC-DS is not important. It's just a way to see Tajo Maturity as a
SQL engine. TPC-DS does not mean anything to TAJO roadmap.

More specifically, the importance is the same with TPC-H, however Tajo
handled TPC-H well now, doesn't it? So, we need just another testsuite to
see how Tajo handle SQL(including WITH/IN/EXISTS) clause well. I think you
agree that in TAJO-1410.

TPC defines
- TPC-H: TPC-H is an ad-hoc, decision support benchmark.
- TPC-DS: TPC Benchmark™DS (TPC-DS): The New Decision Support Benchmark
Standard

In these days, TPC-DS is replacing the ancient TPC-H from real site
customers.
The main reason is based on the lack of TPC-H. TPC-H became a basic level.

Could you tell me if you have some other testcases in mind, except TPC-DS?

Dongjoon.

On Sun, Mar 22, 2015 at 3:58 AM, Hyunsik Choi <hy...@apache.org> wrote:

> Could you tell me why you think TPC-DS is important?
>
> On Saturday, March 21, 2015, Dongjoon Hyun <do...@apache.org> wrote:
>
> > Oh, that was an expression of my regret about my weak pull-requests.
> There
> > exists no offence. It's a duty. I hope you didn't waste your time much
> due
> > to my stupid learning cost.
> >
> > After Hyunsik's direct advice, I have studied Tajo Jira more and
> understand
> > what you mean slightly more. Now I understand many speedup issues, e.g.
> > offheap-based scanner, zero-copy physical layer processing. For newbies
> > like me, we need more men like you.
> >
> > For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap
> already. I
> > respect the current roadmap. You need not worry about that. It's my
> > personal hope as I mentioned before in this mailing list. Here is more
> > recent another report about TPC-DS. (It's consistent with the report
> shared
> > by Hyunsik.)
> >
> >
> >
> https://developer.ibm.com/hadoop/blog/2014/12/02/big-sql-3-0-hadoop-ds-benchmark-performance-isnt-everything/
> >
> >
> > Thank you all.
> >
> > Warmly,
> > Dongjoon.
> >
>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

Could you tell me why you think TPC-DS is important?

On Saturday, March 21, 2015, Dongjoon Hyun <do...@apache.org> wrote:

> Oh, that was an expression of my regret about my weak pull-requests. There
> exists no offence. It's a duty. I hope you didn't waste your time much due
> to my stupid learning cost.
>
> After Hyunsik's direct advice, I have studied Tajo Jira more and understand
> what you mean slightly more. Now I understand many speedup issues, e.g.
> offheap-based scanner, zero-copy physical layer processing. For newbies
> like me, we need more men like you.
>
> For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap already. I
> respect the current roadmap. You need not worry about that. It's my
> personal hope as I mentioned before in this mailing list. Here is more
> recent another report about TPC-DS. (It's consistent with the report shared
> by Hyunsik.)
>
>
> https://developer.ibm.com/hadoop/blog/2014/12/02/big-sql-3-0-hadoop-ds-benchmark-performance-isnt-everything/
>
>
> Thank you all.
>
> Warmly,
> Dongjoon.
>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Dongjoon Hyun <do...@apache.org>]

Oh, that was an expression of my regret about my weak pull-requests. There
exists no offence. It's a duty. I hope you didn't waste your time much due
to my stupid learning cost.

After Hyunsik's direct advice, I have studied Tajo Jira more and understand
what you mean slightly more. Now I understand many speedup issues, e.g.
offheap-based scanner, zero-copy physical layer processing. For newbies
like me, we need more men like you.

For the TPC-DS, Tajo PMC doesn't make any committment in Roadmap already. I
respect the current roadmap. You need not worry about that. It's my
personal hope as I mentioned before in this mailing list. Here is more
recent another report about TPC-DS. (It's consistent with the report shared
by Hyunsik.)

https://developer.ibm.com/hadoop/blog/2014/12/02/big-sql-3-0-hadoop-ds-benchmark-performance-isnt-everything/


Thank you all.

Warmly,
Dongjoon.

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

> I see your point, High Quality Codes.

I meant no offence.

On Fri, Mar 20, 2015 at 4:20 PM, Dongjoon Hyun <do...@apache.org> wrote:
> I see your point, High Quality Codes.
>
> Thank you, Hyunsik.  :-)
>
> On Sat, Mar 21, 2015 at 7:08 AM, Hyunsik Choi <hy...@apache.org> wrote:
>
>> General ideas that we firstly need are already listed on the roadmap
>> or Jira. Everybody know what we should do for speedup and TPC-DS.
>>
>> So, we need code instead of talk.
>>
>> If you or your company want to make Tajo faster, contribute more high
>> quality codes to Tajo community. It's that all.
>>
>> On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org>
>> wrote:
>> > Hi, all.
>> >
>> > I have a question in these day.
>> > *Can we improve Tajo performance further than now?*
>> > Since Tajo is well architectured, each module can be improved
>> > independently.
>> > From the SQL parser to the network part, there will be many candidates to
>> > improve.
>> >
>> > If you didn't try due to lack of time to implement, you can share your
>> idea.
>> > As a Tajo community member, please feel free to comment anything.
>> > We can discuss and create more issues for that.
>> >
>> > When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries
>> (up
>> > to now 26 queries) and Tajo speed becomes more faster. It depends on our
>> > coding speed. :-)
>> >
>> > Best regards,
>> > Dongjoon.
>> >
>> > PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
>> > Apache Derby and HBase, http://www.splicemachine.com/ )
>>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Dongjoon Hyun <do...@apache.org>]

I see your point, High Quality Codes.

Thank you, Hyunsik.  :-)

On Sat, Mar 21, 2015 at 7:08 AM, Hyunsik Choi <hy...@apache.org> wrote:

> General ideas that we firstly need are already listed on the roadmap
> or Jira. Everybody know what we should do for speedup and TPC-DS.
>
> So, we need code instead of talk.
>
> If you or your company want to make Tajo faster, contribute more high
> quality codes to Tajo community. It's that all.
>
> On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org>
> wrote:
> > Hi, all.
> >
> > I have a question in these day.
> > *Can we improve Tajo performance further than now?*
> > Since Tajo is well architectured, each module can be improved
> > independently.
> > From the SQL parser to the network part, there will be many candidates to
> > improve.
> >
> > If you didn't try due to lack of time to implement, you can share your
> idea.
> > As a Tajo community member, please feel free to comment anything.
> > We can discuss and create more issues for that.
> >
> > When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries
> (up
> > to now 26 queries) and Tajo speed becomes more faster. It depends on our
> > coding speed. :-)
> >
> > Best regards,
> > Dongjoon.
> >
> > PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
> > Apache Derby and HBase, http://www.splicemachine.com/ )
>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

Correct some typos.

s/looks/looks enough/
s/we should firstly/we should firstly do/

On Fri, Mar 20, 2015 at 4:37 PM, Hyunsik Choi <hy...@apache.org> wrote:
> There were some typos in my previous post. Especially, I meant 'more
> and high quality'.
>
> In my opinion, our roadmap look enough for you guys to achieve
> something you want. It takes not short time to achieve them. I think
> that there is no magic. We should firstly what we already know.
>
> If you need more description of some features in our roadmap, please
> let me know. I'll describe them.
>
> On Fri, Mar 20, 2015 at 3:08 PM, Hyunsik Choi <hy...@apache.org> wrote:
>> General ideas that we firstly need are already listed on the roadmap
>> or Jira. Everybody know what we should do for speedup and TPC-DS.
>>
>> So, we need code instead of talk.
>>
>> If you or your company want to make Tajo faster, contribute more high
>> quality codes to Tajo community. It's that all.
>>
>> On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org> wrote:
>>> Hi, all.
>>>
>>> I have a question in these day.
>>> *Can we improve Tajo performance further than now?*
>>> Since Tajo is well architectured, each module can be improved
>>> independently.
>>> From the SQL parser to the network part, there will be many candidates to
>>> improve.
>>>
>>> If you didn't try due to lack of time to implement, you can share your idea.
>>> As a Tajo community member, please feel free to comment anything.
>>> We can discuss and create more issues for that.
>>>
>>> When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up
>>> to now 26 queries) and Tajo speed becomes more faster. It depends on our
>>> coding speed. :-)
>>>
>>> Best regards,
>>> Dongjoon.
>>>
>>> PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
>>> Apache Derby and HBase, http://www.splicemachine.com/ )

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

There were some typos in my previous post. Especially, I meant 'more
and high quality'.

In my opinion, our roadmap look enough for you guys to achieve
something you want. It takes not short time to achieve them. I think
that there is no magic. We should firstly what we already know.

If you need more description of some features in our roadmap, please
let me know. I'll describe them.

On Fri, Mar 20, 2015 at 3:08 PM, Hyunsik Choi <hy...@apache.org> wrote:
> General ideas that we firstly need are already listed on the roadmap
> or Jira. Everybody know what we should do for speedup and TPC-DS.
>
> So, we need code instead of talk.
>
> If you or your company want to make Tajo faster, contribute more high
> quality codes to Tajo community. It's that all.
>
> On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org> wrote:
>> Hi, all.
>>
>> I have a question in these day.
>> *Can we improve Tajo performance further than now?*
>> Since Tajo is well architectured, each module can be improved
>> independently.
>> From the SQL parser to the network part, there will be many candidates to
>> improve.
>>
>> If you didn't try due to lack of time to implement, you can share your idea.
>> As a Tajo community member, please feel free to comment anything.
>> We can discuss and create more issues for that.
>>
>> When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up
>> to now 26 queries) and Tajo speed becomes more faster. It depends on our
>> coding speed. :-)
>>
>> Best regards,
>> Dongjoon.
>>
>> PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
>> Apache Derby and HBase, http://www.splicemachine.com/ )

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Jihun Kang <yk...@gmail.com>]

Hyunsik's opinion looks better for me. If anyone want to improve the
performance of Tajo, it could be good to post code for this improvement, or
create a jira issue with detailed test results. It is an interesting issue
for me, however, I don't have multi-node tajo cluster and I could not
validate the performance test.

2015-03-21 7:08 GMT+09:00 Hyunsik Choi <hy...@apache.org>:

> General ideas that we firstly need are already listed on the roadmap
> or Jira. Everybody know what we should do for speedup and TPC-DS.
>
> So, we need code instead of talk.
>
> If you or your company want to make Tajo faster, contribute more high
> quality codes to Tajo community. It's that all.
>
> On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org>
> wrote:
> > Hi, all.
> >
> > I have a question in these day.
> > *Can we improve Tajo performance further than now?*
> > Since Tajo is well architectured, each module can be improved
> > independently.
> > From the SQL parser to the network part, there will be many candidates to
> > improve.
> >
> > If you didn't try due to lack of time to implement, you can share your
> idea.
> > As a Tajo community member, please feel free to comment anything.
> > We can discuss and create more issues for that.
> >
> > When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries
> (up
> > to now 26 queries) and Tajo speed becomes more faster. It depends on our
> > coding speed. :-)
> >
> > Best regards,
> > Dongjoon.
> >
> > PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
> > Apache Derby and HBase, http://www.splicemachine.com/ )
>

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

General ideas that we firstly need are already listed on the roadmap
or Jira. Everybody know what we should do for speedup and TPC-DS.

So, we need code instead of talk.

If you or your company want to make Tajo faster, contribute more high
quality codes to Tajo community. It's that all.

On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org> wrote:
> Hi, all.
>
> I have a question in these day.
> *Can we improve Tajo performance further than now?*
> Since Tajo is well architectured, each module can be improved
> independently.
> From the SQL parser to the network part, there will be many candidates to
> improve.
>
> If you didn't try due to lack of time to implement, you can share your idea.
> As a Tajo community member, please feel free to comment anything.
> We can discuss and create more issues for that.
>
> When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up
> to now 26 queries) and Tajo speed becomes more faster. It depends on our
> coding speed. :-)
>
> Best regards,
> Dongjoon.
>
> PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
> Apache Derby and HBase, http://www.splicemachine.com/ )

Re: [DISCUSSION] Hi-Speed Tajo: Fastest Hadoop DW

[Hyunsik Choi <hy...@apache.org>]

> When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up to now 26 queries) and Tajo speed becomes more faster. It depends on our coding speed. :-)

I heard that they are 26 unmodified queries. If so, it's very impressive.

Please refer to this paper work. According to this paper, impala can
covers about 30, Hive can cover 19, and Presto can only cover 12
original queries.

http://blog.pivotal.io/pivotal/products/pivotal-hawq-benchmark-demonstrates-up-to-21x-faster-performance-on-hadoop-queries-than-sql-like-solutions

It's reality. Of course, we will try to improve Tajo to cover more
TPC-DS benchmarks. I really want too. But, I'm not sure if 0.11.0
release can cover all unmodified TPC-DS queries. If we have full
coverage of TPC-DS as a release condition, the release will be very
delayed.

Best regards,
Hyunsik

On Fri, Mar 20, 2015 at 9:33 AM, Dongjoon Hyun <do...@apache.org> wrote:
> Hi, all.
>
> I have a question in these day.
> *Can we improve Tajo performance further than now?*
> Since Tajo is well architectured, each module can be improved
> independently.
> From the SQL parser to the network part, there will be many candidates to
> improve.
>
> If you didn't try due to lack of time to implement, you can share your idea.
> As a Tajo community member, please feel free to comment anything.
> We can discuss and create more issues for that.
>
> When we release Tajo 0.11, I believe that Tajo passes TPC-DS 99 queries (up
> to now 26 queries) and Tajo speed becomes more faster. It depends on our
> coding speed. :-)
>
> Best regards,
> Dongjoon.
>
> PS. FYI, 'The Only Hadoop RDBMS' is used for SpliceMachine. (based on
> Apache Derby and HBase, http://www.splicemachine.com/ )