quickstep-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Julian Hyde <jh...@apache.org>
Subject Re: [jira] [Commented] (QUICKSTEP-20) Add parser support for SQL window aggregation function
Date Wed, 15 Jun 2016 20:45:19 GMT
If you’re going to build a full database, go for it, best of luck to you, but you’re going
to need a large community.

But in my opinion, you should focus on your strengths. The community should ask itself what
is the one thing that it can do better than anyone else. If you build library that can be
embedded in other projects and in commercial products, you will gain adoption. If you try
to solve the whole problem, everyone (including other Apache projects) will be against you.

Research projects and companies have an interest in “going big”. But projects, in my experience,
work better if you go small.

That said, it does help to build community if you allow people to contribute stuff that interests
them. But make sure you know what your goal is, and how those contributions contribute to
that goal.

Julian


> On Jun 15, 2016, at 8:15 AM, Jignesh Patel <jipatel@pivotal.io> wrote:
> 
> Great points Julian, especially about algebra. Couldn’t agree more.
> 
> In fact, we have been strong advocates of the viewpoint that it is all about the algebraic
framework. Furthermore, we have argued that the relational algebraic framework is the right
“core” to build a platform. With it you can go well beyond warehousing/SQL but also (with
small extensions) build:
> 
> #1: JSON document stores (see Argo <http://pages.cs.wisc.edu/~chasseur/pubs/argo-short.pdf>),

> 
> #2: Iterative graph analytics (see Grail <http://www.cs.wisc.edu/~jignesh/publ/Grail.pdf>),

> 
> #3: Relational learning (see QuickFOIL <http://www.cs.wisc.edu/~jignesh/publ/QuickFoil.pdf>),
and
> 
> #4: Biological data management (see Periscope/SQ <http://www.vldb.org/conf/2007/papers/demo/p1406-tata.pdf>
and Periscope/GQ <http://www.vldb.org/pvldb/1/1454184.pdf>).
> 
> If all of that is not enough, there are nice synergies between deeper integration of
common classes of machine learning and relational data representation. A key idea here is
factorized learning, which my student Arun Kumar (co-advised with Naughton) introduced last
year <http://pages.cs.wisc.edu/~arun/orion/LearningOverJoinsSIGMOD.pdf>. Arun will present
a far deeper follow-on paper <http://pages.cs.wisc.edu/~arun/hamlet/OptFSSIGMOD.pdf>
on this topic at SIGMOD in a few weeks. Interestingly, many other papers are starting to build
on these initial ideas. There is still a bunch of theory to figure out, as a research community,
we are collectively getting very close to nailing that.
> 
> In my keynote @ SIGMOD last year <http://dl.acm.org/citation.cfm?doid=2723372.2723374>,
I talked about how theory (see papers above) has shown that with an extended relational algebraic
core these seemingly different applications converge to a platform that is powered by a relational
core. This converged platform is the long-term vision for Quickstep. Yup — I hear you, I
need to write this up for the community. You are right and I’m adding it to my list :-)

> 
> We have shown prototypes for all of the above, but haven’t put it all together. That
is the hard part, and we are at the start of that journey. That effort is also revealing all
kinds of interesting systems research issues — so good for the students on the project.
Potentially exciting times ahead!
> 
> Cheers,
> Jignesh 
> 
>> On Jun 14, 2016, at 2:32 PM, Julian Hyde <jhyde@apache.org> wrote:
>> 
>> Having that representation reduces coupling in your architecture, so is useful even
if you don’t decide to use a library for SQL parsing/planning. But I think once you have
it you will realize that all of the interesting problems for the project happen after the
query has been converted to algebra.
>> 
>> Julian


Mime
View raw message