You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@hive.apache.org by An...@parc.com on 2015/05/15 21:48:59 UTC
Re: Request for feedback on work intent for non-equijoin support
Hello,
At this point, I have implemented a standalone version of the
1-bucket-theta join algorithm described in the northeastern paper on
Hadoop MR, and would like to start porting it to Hive.
I have been looking at the code and believe that the main goal would be to
implement a new JoinOperator. However, it¹s still not very clear to me how
this class interacts with the rest of the platform (i.e. How it fits in
the overall query processing workflow).
Could someone please provide or point me to a crash course on implementing
a join operator? If nothing else, a list of steps and other classes that I
may have to touch or add would be a very helpful starting point.
Also, I suppose tez is preferred for the implementation, right?
Thanks for your help,
Andrés
On 4/8/15, 2:32 PM, "Thejas Nair" <th...@gmail.com> wrote:
>Yes, the theta join paper in northeastern is a good place to start.
>There is also a presentation from the folks in youtube, which is also
>very useful.
>I had a look at this issue as well earlier, and I had written up a
>rough proposal. I had not organized the document well enough for
>sharing publicly, but in case you find it useful, I have attached it
>to wiki -
>https://cwiki.apache.org/confluence/download/attachments/27362075/theta%20
>join%20proposal%20-%20thejas.pdf?version=1&modificationDate=1428517702954&
>api=v2
>It also includes a list of some of the changes that are needed (it is
>probably not comprehensive enough).
>
>
>On Wed, Apr 8, 2015 at 5:49 AM, <An...@parc.com> wrote:
>> So, I'd like to get started on this. The description in the design doc
>>and the theta join paper from Northeastern seem like a good place to
>>start, to have a baseline that I can later use for the more specific
>>join algorithms I want to try.
>>
>> I created a JIRA account, and my username is Andres.Quiroz
>>
>> Brock, since I'm completely new to this code, could you (or anyone
>>else) please point me to the relevant modules to start learning and
>>ramping up? Also, please let me know if I can contact you directly for
>>discussing this specific topic, or if I should always send a message to
>>the mailing list.
>>
>> Thank you,
>>
>> Andrés
>>
>> -----Original Message-----
>> From: Andres.Quiroz@parc.com [mailto:Andres.Quiroz@parc.com]
>> Sent: Thursday, April 02, 2015 9:07 AM
>> To: dev@hive.apache.org
>> Subject: RE: Request for feedback on work intent for non-equijoin
>>support
>>
>> This is a great pointer, Szehon and Brock, thank you. I will catch up
>>with the material on theta joins and circle back.
>>
>> Andrés
>>
>> -----Original Message-----
>> From: Brock Noland [mailto:brock@apache.org]
>> Sent: Thursday, April 02, 2015 1:31 AM
>> To: dev@hive.apache.org
>> Subject: Re: Request for feedback on work intent for non-equijoin
>>support
>>
>> Nice, it'd be great if someone finally implemented this :)
>>
>> On Wed, Apr 1, 2015 at 10:10 PM, Szehon Ho <sz...@cloudera.com> wrote:
>>> From Hive side, there has been some thought on the subject here:
>>> https://cwiki.apache.org/confluence/display/Hive/Theta+Join, it has
>>> some ideas but nobody has gotten around to giving it a try. It might
>>> be of interest.
>>>
>>> Thanks
>>> Szehon
>>>
>>>
>>> On Wed, Apr 1, 2015 at 10:05 PM, Lefty Leverenz
>>> <le...@gmail.com>
>>> wrote:
>>>
>>>> D'oh! Thanks Chao.
>>>>
>>>> -- Lefty
>>>>
>>>> On Thu, Apr 2, 2015 at 12:59 AM, Chao Sun <ch...@cloudera.com> wrote:
>>>>
>>>> > Hey Lefty,
>>>> >
>>>> > You need to use the ftp protocol, not http.
>>>> > After clicking the link, you'll need to remove "http://" from the
>>>> address
>>>> > bar.
>>>> >
>>>> > Best,
>>>> > Chao
>>>> >
>>>> > On Wed, Apr 1, 2015 at 9:41 PM, Lefty Leverenz
>>>> > <le...@gmail.com>
>>>> > wrote:
>>>> >
>>>> > > Andrés, I followed that link and got the dread 404 Not Found:
>>>> > >
>>>> > > "The requested URI /pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>> > > was not found on this server."
>>>> > >
>>>> > > -- Lefty
>>>> > >
>>>> > > On Wed, Apr 1, 2015 at 7:23 PM, <An...@parc.com> wrote:
>>>> > >
>>>> > > > Dear Lefty,
>>>> > > >
>>>> > > > Thank you very much for pointing that out and for your initial
>>>> > pointers.
>>>> > > > Here is the missing link:
>>>> > > >
>>>> > > > ftp.parc.com/pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>> > > >
>>>> > > > Regards,
>>>> > > >
>>>> > > > Andrés
>>>> > > >
>>>> > > > -----Original Message-----
>>>> > > > From: Lefty Leverenz [mailto:leftyleverenz@gmail.com]
>>>> > > > Sent: Wednesday, April 01, 2015 12:48 AM
>>>> > > > To: dev@hive.apache.org
>>>> > > > Subject: Re: Request for feedback on work intent for
>>>> > > > non-equijoin
>>>> > support
>>>> > > >
>>>> > > > Hello Andres, the link to your paper is missing:
>>>> > > >
>>>> > > > In our preliminary work, which you can find here (pointer to
>>>> > > > the
>>>> paper)
>>>> > > ...
>>>> > > >
>>>> > > >
>>>> > > > You can find general information about contributing to Hive in
>>>> > > > the
>>>> > > > wiki: Resources
>>>> > > > for Contributors
>>>> > > > <
>>>> > > >
>>>> > >
>>>> >
>>>> https://cwiki.apache.org/confluence/display/Hive/Home#Home-Resourcesf
>>>> orContributors
>>>> > > > >
>>>> > > > , How to Contribute
>>>> > > >
>>>><https://cwiki.apache.org/confluence/display/Hive/HowToContribute>.
>>>> > > >
>>>> > > > -- Lefty
>>>> > > >
>>>> > > > On Tue, Mar 31, 2015 at 10:42 PM, <An...@parc.com>
>>>>wrote:
>>>> > > >
>>>> > > > > Dear Hive development community members,
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > I am interested in learning more about the current support
>>>> > > > > for non-equijoins in Hive and/or other Hadoop SQL engines,
>>>> > > > > and in
>>>> getting
>>>> > > > > feedback about community interest in more extensive support
>>>> > > > > for
>>>> such
>>>> > a
>>>> > > > > feature. I intend to work on this challenge, assuming people
>>>> > > > > find
>>>> it
>>>> > > > > compelling, and I intend to contribute results to the
>>>>community.
>>>> > Where
>>>> > > > > possible, it would be great to receive feedback and engage in
>>>> > > > > collaborations along the way (for a bit more context, see the
>>>> > > > > postscript of this message).
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > My initial goal is to support query conditions such as the
>>>> following:
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > A.x < B.y
>>>> > > > >
>>>> > > > > A.x in_range [B.y, B.z]
>>>> > > > >
>>>> > > > > distance(A.x, B.y) < D
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > where A and B are distinct tables/files. It is my
>>>> > > > > understanding
>>>> that
>>>> > > > > current support for performing non-equijoins like those above
>>>> > > > > is
>>>> > quite
>>>> > > > > limited, and where some forms are supported (like in
>>>> > > > > Cloudera's Impala), this support is based on doing a
>>>> > > > > potentially expensive
>>>> cross
>>>> > > > product join.
>>>> > > > > Depending on the data types involved, I believe that joins
>>>> > > > > with
>>>> these
>>>> > > > > conditions can be made to be tractable (at least on the
>>>> > > > > average)
>>>> with
>>>> > > > > join algorithms that exploit properties of the data types,
>>>> > > > > possibly with some pre-scanning of the data.
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > I am asking for feedback on the interest & need in the
>>>> > > > > community
>>>> for
>>>> > > > > this work, as well as any pointers to similar work. In
>>>> > > > > particular,
>>>> I
>>>> > > > > would appreciate any answers people could give on the
>>>> > > > > following
>>>> > > > questions:
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > - Is my understanding of the state of the art in Hive and
>>>> > > > > similar tools accurate? Are there groups currently working on
>>>> > > > > similar or related issues, or tools that already accomplish
>>>> > > > > some or all of
>>>> what
>>>> > I
>>>> > > > have proposed?
>>>> > > > >
>>>> > > > > - Is there significant value to the community in the support
>>>> > > > > of
>>>> such
>>>> > a
>>>> > > > > feature? In other words, are the manual workarounds necessary
>>>> because
>>>> > > > > of the absence of non-equijoins such as these enough of a
>>>> > > > > pain to justify the work I propose?
>>>> > > > >
>>>> > > > > - Being aware that the potential pre-scanning adds to the
>>>> > > > > cost of
>>>> the
>>>> > > > > join, and that data could still blow-up in the worst case, am
>>>> > > > > I missing any other important considerations and tradeoffs
>>>> > > > > for this
>>>> > > > problem?
>>>> > > > >
>>>> > > > > - What would be a good avenue to contribute this feature to
>>>> > > > > the community (e.g. as a standalone tool on top of Hadoop, or
>>>> > > > > as a Hive extension or plugin)?
>>>> > > > >
>>>> > > > > - What is the best way to get started in working with the
>>>> community?
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > Thanks for your attention and any info you can provide!
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > Andres Quiroz
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > P.S. If you are interested in some context, and why/how I am
>>>> > proposing
>>>> > > > > to do this work, please read on.
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > I am part of a small project team at PARC working on the
>>>> > > > > general problems of data integration and automated ETL. We
>>>> > > > > have proposed a tool called HiperFuse that is designed to
>>>> > > > > accept declarative, high-level queries in order to produce
>>>> > > > > joined (fused) data sets
>>>> from
>>>> > > > > multiple heterogeneous raw data sources. In our preliminary
>>>> > > > > work, which you can find here (pointer to the paper), we
>>>> > > > > designed the architecture of the tool and obtained some
>>>> > > > > results separately on
>>>> the
>>>> > > > > problems of automated data cleansing, data type inference,
>>>> > > > > and
>>>> query
>>>> > > > > planning. One of the planned prototype implementations of
>>>> > > > > HiperFuse relies on Hadoop MR, and because the declarative
>>>> > > > > language we
>>>> proposed
>>>> > > > > was closely related to SQL, we thought that we could exploit
>>>> > > > > the existing work in Hive and/or other open-source tools for
>>>> > > > > handling
>>>> the
>>>> > > > > SQL part and layer our work on top of that. For example, the
>>>> > > > > query given in the paper could easily be expressed in
>>>> > > > > SQL-like form with
>>>> a
>>>> > > > > non-equijoin
>>>> > > > > condition:
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > SELECT web_access_log.ip, census.income
>>>> > > > >
>>>> > > > > FROM web_access_log, ip2zip, census
>>>> > > > >
>>>> > > > > WHERE web_access_log.ip in_range [ip2zip.ip_low,
>>>> > > > > ip2zip.ip_high]
>>>> > > > >
>>>> > > > > AND ip2zip.zip = census.zip
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > As you can see, the first impasse that we hit in order to
>>>> > > > > bring the elements together to solve this query end-to-end
>>>> > > > > was the
>>>> realization
>>>> > > > > and performance of the non-equality join in the query. The
>>>> > > > > intent
>>>> now
>>>> > > > > is to tackle this problem in a general sense and provide a
>>>> > > > > solution for a wide range of queries.
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > The work I propose to do would be based on three main
>>>> > > > > components within
>>>> > > > > HiperFuse:
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > > > - Enhancements to the extensible data type framework in
>>>> > > > > HiperFuse
>>>> > that
>>>> > > > > would categorize data types based on the properties needed to
>>>> support
>>>> > > > > the join algorithms, in order to write join-ready
>>>> > > > > domain-specific
>>>> > data
>>>> > > > > type libraries.
>>>> > > > >
>>>> > > > > - The join algorithms themselves, based on Hive or directly
>>>> > > > > on
>>>> Hadoop
>>>> > > MR.
>>>> > > > >
>>>> > > > > - A query planner, which would determine the right algorithm
>>>> > > > > to
>>>> apply
>>>> > > > > and automatically schedule any necessary pre-scanning of the
>>>>data.
>>>> > > > >
>>>> > > > >
>>>> > > > >
>>>> > > >
>>>> > >
>>>> >
>>>> >
>>>> >
>>>> > --
>>>> > Best,
>>>> > Chao
>>>> >
>>>>
Re: Request for feedback on work intent for non-equijoin support
Posted by An...@parc.com.
Ok, that would be great! Except for Monday and Friday, I could meet any
day next week in the afternoon (Pacific time), since it is the end of the
day for me.
Thanks a lot,
Andrés
On 5/15/15, 4:13 PM, "Thejas Nair" <th...@gmail.com> wrote:
>Hi Andres,
>Glad to hear about the progress!
>
>Vikram is a hive join implementation expert. He can guide you through
>this.
>We can setup a webex or google hangout and discuss this. Does sometime
>next week work for you ? (Please let us know some hours that work for
>you, in Pacific time zone).
>
>Anybody else who is interested in the theta join work is also welcome
>to join the discussion. Please let me know.
>
>Thanks,
>Thejas
>
>
>On Fri, May 15, 2015 at 12:48 PM, <An...@parc.com> wrote:
>> Hello,
>>
>> At this point, I have implemented a standalone version of the
>> 1-bucket-theta join algorithm described in the northeastern paper on
>> Hadoop MR, and would like to start porting it to Hive.
>>
>> I have been looking at the code and believe that the main goal would be
>>to
>> implement a new JoinOperator. However, it¹s still not very clear to me
>>how
>> this class interacts with the rest of the platform (i.e. How it fits in
>> the overall query processing workflow).
>>
>> Could someone please provide or point me to a crash course on
>>implementing
>> a join operator? If nothing else, a list of steps and other classes
>>that I
>> may have to touch or add would be a very helpful starting point.
>>
>> Also, I suppose tez is preferred for the implementation, right?
>>
>> Thanks for your help,
>>
>> Andrés
>>
>> On 4/8/15, 2:32 PM, "Thejas Nair" <th...@gmail.com> wrote:
>>
>>>Yes, the theta join paper in northeastern is a good place to start.
>>>There is also a presentation from the folks in youtube, which is also
>>>very useful.
>>>I had a look at this issue as well earlier, and I had written up a
>>>rough proposal. I had not organized the document well enough for
>>>sharing publicly, but in case you find it useful, I have attached it
>>>to wiki -
>>>https://cwiki.apache.org/confluence/download/attachments/27362075/theta%
>>>20
>>>join%20proposal%20-%20thejas.pdf?version=1&modificationDate=142851770295
>>>4&
>>>api=v2
>>>It also includes a list of some of the changes that are needed (it is
>>>probably not comprehensive enough).
>>>
>>>
>>>On Wed, Apr 8, 2015 at 5:49 AM, <An...@parc.com> wrote:
>>>> So, I'd like to get started on this. The description in the design doc
>>>>and the theta join paper from Northeastern seem like a good place to
>>>>start, to have a baseline that I can later use for the more specific
>>>>join algorithms I want to try.
>>>>
>>>> I created a JIRA account, and my username is Andres.Quiroz
>>>>
>>>> Brock, since I'm completely new to this code, could you (or anyone
>>>>else) please point me to the relevant modules to start learning and
>>>>ramping up? Also, please let me know if I can contact you directly for
>>>>discussing this specific topic, or if I should always send a message to
>>>>the mailing list.
>>>>
>>>> Thank you,
>>>>
>>>> Andrés
>>>>
>>>> -----Original Message-----
>>>> From: Andres.Quiroz@parc.com [mailto:Andres.Quiroz@parc.com]
>>>> Sent: Thursday, April 02, 2015 9:07 AM
>>>> To: dev@hive.apache.org
>>>> Subject: RE: Request for feedback on work intent for non-equijoin
>>>>support
>>>>
>>>> This is a great pointer, Szehon and Brock, thank you. I will catch up
>>>>with the material on theta joins and circle back.
>>>>
>>>> Andrés
>>>>
>>>> -----Original Message-----
>>>> From: Brock Noland [mailto:brock@apache.org]
>>>> Sent: Thursday, April 02, 2015 1:31 AM
>>>> To: dev@hive.apache.org
>>>> Subject: Re: Request for feedback on work intent for non-equijoin
>>>>support
>>>>
>>>> Nice, it'd be great if someone finally implemented this :)
>>>>
>>>> On Wed, Apr 1, 2015 at 10:10 PM, Szehon Ho <sz...@cloudera.com>
>>>>wrote:
>>>>> From Hive side, there has been some thought on the subject here:
>>>>> https://cwiki.apache.org/confluence/display/Hive/Theta+Join, it has
>>>>> some ideas but nobody has gotten around to giving it a try. It might
>>>>> be of interest.
>>>>>
>>>>> Thanks
>>>>> Szehon
>>>>>
>>>>>
>>>>> On Wed, Apr 1, 2015 at 10:05 PM, Lefty Leverenz
>>>>> <le...@gmail.com>
>>>>> wrote:
>>>>>
>>>>>> D'oh! Thanks Chao.
>>>>>>
>>>>>> -- Lefty
>>>>>>
>>>>>> On Thu, Apr 2, 2015 at 12:59 AM, Chao Sun <ch...@cloudera.com> wrote:
>>>>>>
>>>>>> > Hey Lefty,
>>>>>> >
>>>>>> > You need to use the ftp protocol, not http.
>>>>>> > After clicking the link, you'll need to remove "http://" from the
>>>>>> address
>>>>>> > bar.
>>>>>> >
>>>>>> > Best,
>>>>>> > Chao
>>>>>> >
>>>>>> > On Wed, Apr 1, 2015 at 9:41 PM, Lefty Leverenz
>>>>>> > <le...@gmail.com>
>>>>>> > wrote:
>>>>>> >
>>>>>> > > Andrés, I followed that link and got the dread 404 Not Found:
>>>>>> > >
>>>>>> > > "The requested URI /pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>>>> > > was not found on this server."
>>>>>> > >
>>>>>> > > -- Lefty
>>>>>> > >
>>>>>> > > On Wed, Apr 1, 2015 at 7:23 PM, <An...@parc.com> wrote:
>>>>>> > >
>>>>>> > > > Dear Lefty,
>>>>>> > > >
>>>>>> > > > Thank you very much for pointing that out and for your initial
>>>>>> > pointers.
>>>>>> > > > Here is the missing link:
>>>>>> > > >
>>>>>> > > > ftp.parc.com/pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>>>> > > >
>>>>>> > > > Regards,
>>>>>> > > >
>>>>>> > > > Andrés
>>>>>> > > >
>>>>>> > > > -----Original Message-----
>>>>>> > > > From: Lefty Leverenz [mailto:leftyleverenz@gmail.com]
>>>>>> > > > Sent: Wednesday, April 01, 2015 12:48 AM
>>>>>> > > > To: dev@hive.apache.org
>>>>>> > > > Subject: Re: Request for feedback on work intent for
>>>>>> > > > non-equijoin
>>>>>> > support
>>>>>> > > >
>>>>>> > > > Hello Andres, the link to your paper is missing:
>>>>>> > > >
>>>>>> > > > In our preliminary work, which you can find here (pointer to
>>>>>> > > > the
>>>>>> paper)
>>>>>> > > ...
>>>>>> > > >
>>>>>> > > >
>>>>>> > > > You can find general information about contributing to Hive in
>>>>>> > > > the
>>>>>> > > > wiki: Resources
>>>>>> > > > for Contributors
>>>>>> > > > <
>>>>>> > > >
>>>>>> > >
>>>>>> >
>>>>>>
>>>>>>https://cwiki.apache.org/confluence/display/Hive/Home#Home-Resourcesf
>>>>>> orContributors
>>>>>> > > > >
>>>>>> > > > , How to Contribute
>>>>>> > > >
>>>>>><https://cwiki.apache.org/confluence/display/Hive/HowToContribute>.
>>>>>> > > >
>>>>>> > > > -- Lefty
>>>>>> > > >
>>>>>> > > > On Tue, Mar 31, 2015 at 10:42 PM, <An...@parc.com>
>>>>>>wrote:
>>>>>> > > >
>>>>>> > > > > Dear Hive development community members,
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > I am interested in learning more about the current support
>>>>>> > > > > for non-equijoins in Hive and/or other Hadoop SQL engines,
>>>>>> > > > > and in
>>>>>> getting
>>>>>> > > > > feedback about community interest in more extensive support
>>>>>> > > > > for
>>>>>> such
>>>>>> > a
>>>>>> > > > > feature. I intend to work on this challenge, assuming people
>>>>>> > > > > find
>>>>>> it
>>>>>> > > > > compelling, and I intend to contribute results to the
>>>>>>community.
>>>>>> > Where
>>>>>> > > > > possible, it would be great to receive feedback and engage
>>>>>>in
>>>>>> > > > > collaborations along the way (for a bit more context, see
>>>>>>the
>>>>>> > > > > postscript of this message).
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > My initial goal is to support query conditions such as the
>>>>>> following:
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > A.x < B.y
>>>>>> > > > >
>>>>>> > > > > A.x in_range [B.y, B.z]
>>>>>> > > > >
>>>>>> > > > > distance(A.x, B.y) < D
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > where A and B are distinct tables/files. It is my
>>>>>> > > > > understanding
>>>>>> that
>>>>>> > > > > current support for performing non-equijoins like those
>>>>>>above
>>>>>> > > > > is
>>>>>> > quite
>>>>>> > > > > limited, and where some forms are supported (like in
>>>>>> > > > > Cloudera's Impala), this support is based on doing a
>>>>>> > > > > potentially expensive
>>>>>> cross
>>>>>> > > > product join.
>>>>>> > > > > Depending on the data types involved, I believe that joins
>>>>>> > > > > with
>>>>>> these
>>>>>> > > > > conditions can be made to be tractable (at least on the
>>>>>> > > > > average)
>>>>>> with
>>>>>> > > > > join algorithms that exploit properties of the data types,
>>>>>> > > > > possibly with some pre-scanning of the data.
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > I am asking for feedback on the interest & need in the
>>>>>> > > > > community
>>>>>> for
>>>>>> > > > > this work, as well as any pointers to similar work. In
>>>>>> > > > > particular,
>>>>>> I
>>>>>> > > > > would appreciate any answers people could give on the
>>>>>> > > > > following
>>>>>> > > > questions:
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > - Is my understanding of the state of the art in Hive and
>>>>>> > > > > similar tools accurate? Are there groups currently working
>>>>>>on
>>>>>> > > > > similar or related issues, or tools that already accomplish
>>>>>> > > > > some or all of
>>>>>> what
>>>>>> > I
>>>>>> > > > have proposed?
>>>>>> > > > >
>>>>>> > > > > - Is there significant value to the community in the support
>>>>>> > > > > of
>>>>>> such
>>>>>> > a
>>>>>> > > > > feature? In other words, are the manual workarounds
>>>>>>necessary
>>>>>> because
>>>>>> > > > > of the absence of non-equijoins such as these enough of a
>>>>>> > > > > pain to justify the work I propose?
>>>>>> > > > >
>>>>>> > > > > - Being aware that the potential pre-scanning adds to the
>>>>>> > > > > cost of
>>>>>> the
>>>>>> > > > > join, and that data could still blow-up in the worst case,
>>>>>>am
>>>>>> > > > > I missing any other important considerations and tradeoffs
>>>>>> > > > > for this
>>>>>> > > > problem?
>>>>>> > > > >
>>>>>> > > > > - What would be a good avenue to contribute this feature to
>>>>>> > > > > the community (e.g. as a standalone tool on top of Hadoop,
>>>>>>or
>>>>>> > > > > as a Hive extension or plugin)?
>>>>>> > > > >
>>>>>> > > > > - What is the best way to get started in working with the
>>>>>> community?
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > Thanks for your attention and any info you can provide!
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > Andres Quiroz
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > P.S. If you are interested in some context, and why/how I am
>>>>>> > proposing
>>>>>> > > > > to do this work, please read on.
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > I am part of a small project team at PARC working on the
>>>>>> > > > > general problems of data integration and automated ETL. We
>>>>>> > > > > have proposed a tool called HiperFuse that is designed to
>>>>>> > > > > accept declarative, high-level queries in order to produce
>>>>>> > > > > joined (fused) data sets
>>>>>> from
>>>>>> > > > > multiple heterogeneous raw data sources. In our preliminary
>>>>>> > > > > work, which you can find here (pointer to the paper), we
>>>>>> > > > > designed the architecture of the tool and obtained some
>>>>>> > > > > results separately on
>>>>>> the
>>>>>> > > > > problems of automated data cleansing, data type inference,
>>>>>> > > > > and
>>>>>> query
>>>>>> > > > > planning. One of the planned prototype implementations of
>>>>>> > > > > HiperFuse relies on Hadoop MR, and because the declarative
>>>>>> > > > > language we
>>>>>> proposed
>>>>>> > > > > was closely related to SQL, we thought that we could exploit
>>>>>> > > > > the existing work in Hive and/or other open-source tools for
>>>>>> > > > > handling
>>>>>> the
>>>>>> > > > > SQL part and layer our work on top of that. For example, the
>>>>>> > > > > query given in the paper could easily be expressed in
>>>>>> > > > > SQL-like form with
>>>>>> a
>>>>>> > > > > non-equijoin
>>>>>> > > > > condition:
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > SELECT web_access_log.ip, census.income
>>>>>> > > > >
>>>>>> > > > > FROM web_access_log, ip2zip, census
>>>>>> > > > >
>>>>>> > > > > WHERE web_access_log.ip in_range [ip2zip.ip_low,
>>>>>> > > > > ip2zip.ip_high]
>>>>>> > > > >
>>>>>> > > > > AND ip2zip.zip = census.zip
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > As you can see, the first impasse that we hit in order to
>>>>>> > > > > bring the elements together to solve this query end-to-end
>>>>>> > > > > was the
>>>>>> realization
>>>>>> > > > > and performance of the non-equality join in the query. The
>>>>>> > > > > intent
>>>>>> now
>>>>>> > > > > is to tackle this problem in a general sense and provide a
>>>>>> > > > > solution for a wide range of queries.
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > The work I propose to do would be based on three main
>>>>>> > > > > components within
>>>>>> > > > > HiperFuse:
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > > - Enhancements to the extensible data type framework in
>>>>>> > > > > HiperFuse
>>>>>> > that
>>>>>> > > > > would categorize data types based on the properties needed
>>>>>>to
>>>>>> support
>>>>>> > > > > the join algorithms, in order to write join-ready
>>>>>> > > > > domain-specific
>>>>>> > data
>>>>>> > > > > type libraries.
>>>>>> > > > >
>>>>>> > > > > - The join algorithms themselves, based on Hive or directly
>>>>>> > > > > on
>>>>>> Hadoop
>>>>>> > > MR.
>>>>>> > > > >
>>>>>> > > > > - A query planner, which would determine the right algorithm
>>>>>> > > > > to
>>>>>> apply
>>>>>> > > > > and automatically schedule any necessary pre-scanning of the
>>>>>>data.
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > > >
>>>>>> > > >
>>>>>> > >
>>>>>> >
>>>>>> >
>>>>>> >
>>>>>> > --
>>>>>> > Best,
>>>>>> > Chao
>>>>>> >
>>>>>>
>>
Re: Request for feedback on work intent for non-equijoin support
Posted by Thejas Nair <th...@gmail.com>.
Hi Andres,
Glad to hear about the progress!
Vikram is a hive join implementation expert. He can guide you through this.
We can setup a webex or google hangout and discuss this. Does sometime
next week work for you ? (Please let us know some hours that work for
you, in Pacific time zone).
Anybody else who is interested in the theta join work is also welcome
to join the discussion. Please let me know.
Thanks,
Thejas
On Fri, May 15, 2015 at 12:48 PM, <An...@parc.com> wrote:
> Hello,
>
> At this point, I have implemented a standalone version of the
> 1-bucket-theta join algorithm described in the northeastern paper on
> Hadoop MR, and would like to start porting it to Hive.
>
> I have been looking at the code and believe that the main goal would be to
> implement a new JoinOperator. However, it¹s still not very clear to me how
> this class interacts with the rest of the platform (i.e. How it fits in
> the overall query processing workflow).
>
> Could someone please provide or point me to a crash course on implementing
> a join operator? If nothing else, a list of steps and other classes that I
> may have to touch or add would be a very helpful starting point.
>
> Also, I suppose tez is preferred for the implementation, right?
>
> Thanks for your help,
>
> Andrés
>
> On 4/8/15, 2:32 PM, "Thejas Nair" <th...@gmail.com> wrote:
>
>>Yes, the theta join paper in northeastern is a good place to start.
>>There is also a presentation from the folks in youtube, which is also
>>very useful.
>>I had a look at this issue as well earlier, and I had written up a
>>rough proposal. I had not organized the document well enough for
>>sharing publicly, but in case you find it useful, I have attached it
>>to wiki -
>>https://cwiki.apache.org/confluence/download/attachments/27362075/theta%20
>>join%20proposal%20-%20thejas.pdf?version=1&modificationDate=1428517702954&
>>api=v2
>>It also includes a list of some of the changes that are needed (it is
>>probably not comprehensive enough).
>>
>>
>>On Wed, Apr 8, 2015 at 5:49 AM, <An...@parc.com> wrote:
>>> So, I'd like to get started on this. The description in the design doc
>>>and the theta join paper from Northeastern seem like a good place to
>>>start, to have a baseline that I can later use for the more specific
>>>join algorithms I want to try.
>>>
>>> I created a JIRA account, and my username is Andres.Quiroz
>>>
>>> Brock, since I'm completely new to this code, could you (or anyone
>>>else) please point me to the relevant modules to start learning and
>>>ramping up? Also, please let me know if I can contact you directly for
>>>discussing this specific topic, or if I should always send a message to
>>>the mailing list.
>>>
>>> Thank you,
>>>
>>> Andrés
>>>
>>> -----Original Message-----
>>> From: Andres.Quiroz@parc.com [mailto:Andres.Quiroz@parc.com]
>>> Sent: Thursday, April 02, 2015 9:07 AM
>>> To: dev@hive.apache.org
>>> Subject: RE: Request for feedback on work intent for non-equijoin
>>>support
>>>
>>> This is a great pointer, Szehon and Brock, thank you. I will catch up
>>>with the material on theta joins and circle back.
>>>
>>> Andrés
>>>
>>> -----Original Message-----
>>> From: Brock Noland [mailto:brock@apache.org]
>>> Sent: Thursday, April 02, 2015 1:31 AM
>>> To: dev@hive.apache.org
>>> Subject: Re: Request for feedback on work intent for non-equijoin
>>>support
>>>
>>> Nice, it'd be great if someone finally implemented this :)
>>>
>>> On Wed, Apr 1, 2015 at 10:10 PM, Szehon Ho <sz...@cloudera.com> wrote:
>>>> From Hive side, there has been some thought on the subject here:
>>>> https://cwiki.apache.org/confluence/display/Hive/Theta+Join, it has
>>>> some ideas but nobody has gotten around to giving it a try. It might
>>>> be of interest.
>>>>
>>>> Thanks
>>>> Szehon
>>>>
>>>>
>>>> On Wed, Apr 1, 2015 at 10:05 PM, Lefty Leverenz
>>>> <le...@gmail.com>
>>>> wrote:
>>>>
>>>>> D'oh! Thanks Chao.
>>>>>
>>>>> -- Lefty
>>>>>
>>>>> On Thu, Apr 2, 2015 at 12:59 AM, Chao Sun <ch...@cloudera.com> wrote:
>>>>>
>>>>> > Hey Lefty,
>>>>> >
>>>>> > You need to use the ftp protocol, not http.
>>>>> > After clicking the link, you'll need to remove "http://" from the
>>>>> address
>>>>> > bar.
>>>>> >
>>>>> > Best,
>>>>> > Chao
>>>>> >
>>>>> > On Wed, Apr 1, 2015 at 9:41 PM, Lefty Leverenz
>>>>> > <le...@gmail.com>
>>>>> > wrote:
>>>>> >
>>>>> > > Andrés, I followed that link and got the dread 404 Not Found:
>>>>> > >
>>>>> > > "The requested URI /pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>>> > > was not found on this server."
>>>>> > >
>>>>> > > -- Lefty
>>>>> > >
>>>>> > > On Wed, Apr 1, 2015 at 7:23 PM, <An...@parc.com> wrote:
>>>>> > >
>>>>> > > > Dear Lefty,
>>>>> > > >
>>>>> > > > Thank you very much for pointing that out and for your initial
>>>>> > pointers.
>>>>> > > > Here is the missing link:
>>>>> > > >
>>>>> > > > ftp.parc.com/pub/torres/Hiperfuse/extended_hiperfuse.pdf
>>>>> > > >
>>>>> > > > Regards,
>>>>> > > >
>>>>> > > > Andrés
>>>>> > > >
>>>>> > > > -----Original Message-----
>>>>> > > > From: Lefty Leverenz [mailto:leftyleverenz@gmail.com]
>>>>> > > > Sent: Wednesday, April 01, 2015 12:48 AM
>>>>> > > > To: dev@hive.apache.org
>>>>> > > > Subject: Re: Request for feedback on work intent for
>>>>> > > > non-equijoin
>>>>> > support
>>>>> > > >
>>>>> > > > Hello Andres, the link to your paper is missing:
>>>>> > > >
>>>>> > > > In our preliminary work, which you can find here (pointer to
>>>>> > > > the
>>>>> paper)
>>>>> > > ...
>>>>> > > >
>>>>> > > >
>>>>> > > > You can find general information about contributing to Hive in
>>>>> > > > the
>>>>> > > > wiki: Resources
>>>>> > > > for Contributors
>>>>> > > > <
>>>>> > > >
>>>>> > >
>>>>> >
>>>>> https://cwiki.apache.org/confluence/display/Hive/Home#Home-Resourcesf
>>>>> orContributors
>>>>> > > > >
>>>>> > > > , How to Contribute
>>>>> > > >
>>>>><https://cwiki.apache.org/confluence/display/Hive/HowToContribute>.
>>>>> > > >
>>>>> > > > -- Lefty
>>>>> > > >
>>>>> > > > On Tue, Mar 31, 2015 at 10:42 PM, <An...@parc.com>
>>>>>wrote:
>>>>> > > >
>>>>> > > > > Dear Hive development community members,
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > I am interested in learning more about the current support
>>>>> > > > > for non-equijoins in Hive and/or other Hadoop SQL engines,
>>>>> > > > > and in
>>>>> getting
>>>>> > > > > feedback about community interest in more extensive support
>>>>> > > > > for
>>>>> such
>>>>> > a
>>>>> > > > > feature. I intend to work on this challenge, assuming people
>>>>> > > > > find
>>>>> it
>>>>> > > > > compelling, and I intend to contribute results to the
>>>>>community.
>>>>> > Where
>>>>> > > > > possible, it would be great to receive feedback and engage in
>>>>> > > > > collaborations along the way (for a bit more context, see the
>>>>> > > > > postscript of this message).
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > My initial goal is to support query conditions such as the
>>>>> following:
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > A.x < B.y
>>>>> > > > >
>>>>> > > > > A.x in_range [B.y, B.z]
>>>>> > > > >
>>>>> > > > > distance(A.x, B.y) < D
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > where A and B are distinct tables/files. It is my
>>>>> > > > > understanding
>>>>> that
>>>>> > > > > current support for performing non-equijoins like those above
>>>>> > > > > is
>>>>> > quite
>>>>> > > > > limited, and where some forms are supported (like in
>>>>> > > > > Cloudera's Impala), this support is based on doing a
>>>>> > > > > potentially expensive
>>>>> cross
>>>>> > > > product join.
>>>>> > > > > Depending on the data types involved, I believe that joins
>>>>> > > > > with
>>>>> these
>>>>> > > > > conditions can be made to be tractable (at least on the
>>>>> > > > > average)
>>>>> with
>>>>> > > > > join algorithms that exploit properties of the data types,
>>>>> > > > > possibly with some pre-scanning of the data.
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > I am asking for feedback on the interest & need in the
>>>>> > > > > community
>>>>> for
>>>>> > > > > this work, as well as any pointers to similar work. In
>>>>> > > > > particular,
>>>>> I
>>>>> > > > > would appreciate any answers people could give on the
>>>>> > > > > following
>>>>> > > > questions:
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > - Is my understanding of the state of the art in Hive and
>>>>> > > > > similar tools accurate? Are there groups currently working on
>>>>> > > > > similar or related issues, or tools that already accomplish
>>>>> > > > > some or all of
>>>>> what
>>>>> > I
>>>>> > > > have proposed?
>>>>> > > > >
>>>>> > > > > - Is there significant value to the community in the support
>>>>> > > > > of
>>>>> such
>>>>> > a
>>>>> > > > > feature? In other words, are the manual workarounds necessary
>>>>> because
>>>>> > > > > of the absence of non-equijoins such as these enough of a
>>>>> > > > > pain to justify the work I propose?
>>>>> > > > >
>>>>> > > > > - Being aware that the potential pre-scanning adds to the
>>>>> > > > > cost of
>>>>> the
>>>>> > > > > join, and that data could still blow-up in the worst case, am
>>>>> > > > > I missing any other important considerations and tradeoffs
>>>>> > > > > for this
>>>>> > > > problem?
>>>>> > > > >
>>>>> > > > > - What would be a good avenue to contribute this feature to
>>>>> > > > > the community (e.g. as a standalone tool on top of Hadoop, or
>>>>> > > > > as a Hive extension or plugin)?
>>>>> > > > >
>>>>> > > > > - What is the best way to get started in working with the
>>>>> community?
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > Thanks for your attention and any info you can provide!
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > Andres Quiroz
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > P.S. If you are interested in some context, and why/how I am
>>>>> > proposing
>>>>> > > > > to do this work, please read on.
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > I am part of a small project team at PARC working on the
>>>>> > > > > general problems of data integration and automated ETL. We
>>>>> > > > > have proposed a tool called HiperFuse that is designed to
>>>>> > > > > accept declarative, high-level queries in order to produce
>>>>> > > > > joined (fused) data sets
>>>>> from
>>>>> > > > > multiple heterogeneous raw data sources. In our preliminary
>>>>> > > > > work, which you can find here (pointer to the paper), we
>>>>> > > > > designed the architecture of the tool and obtained some
>>>>> > > > > results separately on
>>>>> the
>>>>> > > > > problems of automated data cleansing, data type inference,
>>>>> > > > > and
>>>>> query
>>>>> > > > > planning. One of the planned prototype implementations of
>>>>> > > > > HiperFuse relies on Hadoop MR, and because the declarative
>>>>> > > > > language we
>>>>> proposed
>>>>> > > > > was closely related to SQL, we thought that we could exploit
>>>>> > > > > the existing work in Hive and/or other open-source tools for
>>>>> > > > > handling
>>>>> the
>>>>> > > > > SQL part and layer our work on top of that. For example, the
>>>>> > > > > query given in the paper could easily be expressed in
>>>>> > > > > SQL-like form with
>>>>> a
>>>>> > > > > non-equijoin
>>>>> > > > > condition:
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > SELECT web_access_log.ip, census.income
>>>>> > > > >
>>>>> > > > > FROM web_access_log, ip2zip, census
>>>>> > > > >
>>>>> > > > > WHERE web_access_log.ip in_range [ip2zip.ip_low,
>>>>> > > > > ip2zip.ip_high]
>>>>> > > > >
>>>>> > > > > AND ip2zip.zip = census.zip
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > As you can see, the first impasse that we hit in order to
>>>>> > > > > bring the elements together to solve this query end-to-end
>>>>> > > > > was the
>>>>> realization
>>>>> > > > > and performance of the non-equality join in the query. The
>>>>> > > > > intent
>>>>> now
>>>>> > > > > is to tackle this problem in a general sense and provide a
>>>>> > > > > solution for a wide range of queries.
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > The work I propose to do would be based on three main
>>>>> > > > > components within
>>>>> > > > > HiperFuse:
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > > > - Enhancements to the extensible data type framework in
>>>>> > > > > HiperFuse
>>>>> > that
>>>>> > > > > would categorize data types based on the properties needed to
>>>>> support
>>>>> > > > > the join algorithms, in order to write join-ready
>>>>> > > > > domain-specific
>>>>> > data
>>>>> > > > > type libraries.
>>>>> > > > >
>>>>> > > > > - The join algorithms themselves, based on Hive or directly
>>>>> > > > > on
>>>>> Hadoop
>>>>> > > MR.
>>>>> > > > >
>>>>> > > > > - A query planner, which would determine the right algorithm
>>>>> > > > > to
>>>>> apply
>>>>> > > > > and automatically schedule any necessary pre-scanning of the
>>>>>data.
>>>>> > > > >
>>>>> > > > >
>>>>> > > > >
>>>>> > > >
>>>>> > >
>>>>> >
>>>>> >
>>>>> >
>>>>> > --
>>>>> > Best,
>>>>> > Chao
>>>>> >
>>>>>
>