[DISCUSS] Flink ML roadmap

[Stavros Kontopoulos]

(Resending with the appropriate topic)

Hi,

I would like to start a discussion about next steps for Flink ML.
Currently there is a lot of work going on but needs a push forward.

Some topics to discuss:

a) How several features should be planned and get aligned with Flink
releases.
b) Priorities of what should be done.
c) Basic guidelines for code: styleguides, scikit-learn compliance etc
d) Missing features important for the success of the library, next steps
etc...

Thoughts?

Best,
Stavros

[Gábor Hermann]

Hi Stavros,

Thanks for bringing this up.

There have been past [1] and recent [2, 3] discussions about the Flink
libraries, because there are some stalling PRs and overloaded
committers. (Actually, Till is the only committer shepherd of the both
the CEP and ML library, and AFAIK he has a ton of other responsibilities
and work to do.) Thus it's hard to get code reviewed and merged, and
without merged code it's hard to get a committer status, so there are
not many committers who can review e.g. ML algorithm implementations,
and the cycle goes on. Until this is resolved somehow, we should help
the committers by reviewing each-others PRs.

I think prioritizing features (b) is a good way to start. We could
declare most blocking features and concentrate on reviewing and merging
them before moving forward. E.g. the evaluation framework is quite
important for an ML library in my opinion, and has a PR stalling for
long [4].

Regarding c),  there are styleguides generally for contributing to
Flink, so we should follow that. Is there something more ML specific you
think we could follow? We should definitely declare, we follow
scikit-learn and make sure contributions comply to that.

In terms of features (a, d), I think we should first see the bigger
picture. That is, it would be nice to discuss a clearer direction for
Flink ML. I've seen a lot of interest in contributing to Flink ML
lately. I believe we should rethink our goals, to put the contribution
efforts in making a usable and useful library. Are we trying to
implement as many useful algorithms as possible to create a scalable ML
library? That would seem ambitious, and of course there are a lot of
frameworks and libraries that already has something like this as goal
(e.g. Spark MLlib, Mahout). Should we rather create connectors to
existing libraries? Then we cannot really do Flink specific
optimizations. Should we go for online machine learning (as Flink is
concentrating on streaming)? We already have a connector to SAMOA. We
could go on with questions like this. Maybe I'm missing something, but I
haven't seen such directions declared.

Cheers,
Gabor

[1]
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/Opening-a-discussion-on-FlinkML-td10265.html
[2]
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/Flink-CEP-development-is-stalling-td15237.html#a15341
[3]
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/New-Flink-team-member-Kate-Eri-td15349.html
[4] https://github.com/apache/flink/pull/1849

On 2017-02-20 11:43, Stavros Kontopoulos wrote:

> (Resending with the appropriate topic)
>
> Hi,
>
> I would like to start a discussion about next steps for Flink ML.
> Currently there is a lot of work going on but needs a push forward.
>
> Some topics to discuss:
>
> a) How several features should be planned and get aligned with Flink
> releases.
> b) Priorities of what should be done.
> c) Basic guidelines for code: styleguides, scikit-learn compliance etc
> d) Missing features important for the success of the library, next steps
> etc...
>
> Thoughts?
>
> Best,
> Stavros
>

[Theodore Vasiloudis]

Hello all,

thank you for opening this discussion Stavros, note that it's almost
exactly 1 year since I last opened such a topic (linked by Gabor) and the
comments there are still relevant.

I think Gabor described the current state quite well, development in the
libraries is hard without committers dedicated to each project, and as a
result FlinkML and CEP have stalled.

I think it's important to look at why development has stalled as well. As
people have mentioned there's a multitude of ML libraries out there and my
impression was that not many people are looking to use Flink for ML. Lately
that seems to have changed (with some interest shown in the Flink survey as
well).

Gabor makes some good points about future directions for the library. Our
initial goal [1] was to make a truly scalable, easy to use library, within
the Flink ecosystem, providing a set of "workhorse" algorithms, sampled
from what's actually being used in the industry. We planned for a library
that has few algorithms, but does them properly.

If we decide to go the way of focusing within Flink we face some major
challenges, because these are system limitations that do not necessarily
align with the goals of the community. Some issues relevant to ML on Flink
are:

   - FLINK-2396 - Review the datasets of dynamic path and static path in
   iteration.
   https://issues.apache.org/jira/browse/FLINK-2396
   This has to do with the ability to iterate over one datset (model) while
   changing another (dataset), which is necessary for many ML algorithms like
   SGD.
   - FLINK-1730 - Add a FlinkTools.persist style method to the Data Set.
   https://issues.apache.org/jira/browse/FLINK-1730
   This is again relevant to many algorithms, to create intermediate
   results etc, for example L-BFGS development has been attempted 2-3 times,
   but always abandoned because of the need to collect a DataSet kills the
   performance.
   - FLINK-5782 - Support GPU calculations
   https://issues.apache.org/jira/browse/FLINK-5782
   Many algorithms will benefit greatly by GPU-accelerated linear algebra,
   to the point where if a library doesn't support it puts it at a severe
   disadvantage compared to other offerings.


These issues aside, Stephan has mentioned recently the possibility of
re-structuring the Flink project to allow for more flexibility for the
libraries. I think that sounds quite promising and it should allow the
development to pick up in the libraries, if we can get some more people
reviewing and merging PRs.

I would be all for updating our vision and roadmap to match what the
community desires from the library.

[1]
https://cwiki.apache.org/confluence/display/FLINK/FlinkML%3A+Vision+and+Roadmap

On Mon, Feb 20, 2017 at 12:47 PM, Gábor Hermann <[hidden email]>
wrote:

> Hi Stavros,
>
> Thanks for bringing this up.
>
> There have been past [1] and recent [2, 3] discussions about the Flink
> libraries, because there are some stalling PRs and overloaded committers.
> (Actually, Till is the only committer shepherd of the both the CEP and ML
> library, and AFAIK he has a ton of other responsibilities and work to do.)
> Thus it's hard to get code reviewed and merged, and without merged code
> it's hard to get a committer status, so there are not many committers who
> can review e.g. ML algorithm implementations, and the cycle goes on. Until
> this is resolved somehow, we should help the committers by reviewing
> each-others PRs.
>
> I think prioritizing features (b) is a good way to start. We could declare
> most blocking features and concentrate on reviewing and merging them before
> moving forward. E.g. the evaluation framework is quite important for an ML
> library in my opinion, and has a PR stalling for long [4].
>
> Regarding c),  there are styleguides generally for contributing to Flink,
> so we should follow that. Is there something more ML specific you think we
> could follow? We should definitely declare, we follow scikit-learn and make
> sure contributions comply to that.
>
> In terms of features (a, d), I think we should first see the bigger
> picture. That is, it would be nice to discuss a clearer direction for Flink
> ML. I've seen a lot of interest in contributing to Flink ML lately. I
> believe we should rethink our goals, to put the contribution efforts in
> making a usable and useful library. Are we trying to implement as many
> useful algorithms as possible to create a scalable ML library? That would
> seem ambitious, and of course there are a lot of frameworks and libraries
> that already has something like this as goal (e.g. Spark MLlib, Mahout).
> Should we rather create connectors to existing libraries? Then we cannot
> really do Flink specific optimizations. Should we go for online machine
> learning (as Flink is concentrating on streaming)? We already have a
> connector to SAMOA. We could go on with questions like this. Maybe I'm
> missing something, but I haven't seen such directions declared.
>
> Cheers,
> Gabor
>
> [1] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> com/Opening-a-discussion-on-FlinkML-td10265.html
> [2] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> com/Flink-CEP-development-is-stalling-td15237.html#a15341
> [3] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> com/New-Flink-team-member-Kate-Eri-td15349.html
> [4] https://github.com/apache/flink/pull/1849
>
>
> On 2017-02-20 11:43, Stavros Kontopoulos wrote:
>
> (Resending with the appropriate topic)
>>
>> Hi,
>>
>> I would like to start a discussion about next steps for Flink ML.
>> Currently there is a lot of work going on but needs a push forward.
>>
>> Some topics to discuss:
>>
>> a) How several features should be planned and get aligned with Flink
>> releases.
>> b) Priorities of what should be done.
>> c) Basic guidelines for code: styleguides, scikit-learn compliance etc
>> d) Missing features important for the success of the library, next steps
>> etc...
>>
>> Thoughts?
>>
>> Best,
>> Stavros
>>
>>
>

[Katherin Eri]

Hello guys,


May be we will be able to focus our forces on some E2E scenario or show
case for Flink as also ML supporting engine, and in such a way actualize
the roadmap?


This means: we can take some real life/production problem, like Fraud
detection in some area, and try to solve this problem from the point of
view of DataScience.

Starting from data preprocessing and preparation, finishing
implementation/usage of some ML algorithm.

Doing this we will understand which issues are showstopper for
implementation of such functionality. We will be able to understand Flink’s
users better.


May be community could share its ideas which show case could be the most
useful for Apache Flink, or may be Data artisans could lead this?

пн, 20 февр. 2017 г. в 15:28, Theodore Vasiloudis <
[hidden email]>:

> Hello all,
>
> thank you for opening this discussion Stavros, note that it's almost
> exactly 1 year since I last opened such a topic (linked by Gabor) and the
> comments there are still relevant.
>
> I think Gabor described the current state quite well, development in the
> libraries is hard without committers dedicated to each project, and as a
> result FlinkML and CEP have stalled.
>
> I think it's important to look at why development has stalled as well. As
> people have mentioned there's a multitude of ML libraries out there and my
> impression was that not many people are looking to use Flink for ML. Lately
> that seems to have changed (with some interest shown in the Flink survey as
> well).
>
> Gabor makes some good points about future directions for the library. Our
> initial goal [1] was to make a truly scalable, easy to use library, within
> the Flink ecosystem, providing a set of "workhorse" algorithms, sampled
> from what's actually being used in the industry. We planned for a library
> that has few algorithms, but does them properly.
>
> If we decide to go the way of focusing within Flink we face some major
> challenges, because these are system limitations that do not necessarily
> align with the goals of the community. Some issues relevant to ML on Flink
> are:
>
>    - FLINK-2396 - Review the datasets of dynamic path and static path in
>    iteration.
>    https://issues.apache.org/jira/browse/FLINK-2396
>    This has to do with the ability to iterate over one datset (model) while
>    changing another (dataset), which is necessary for many ML algorithms
> like
>    SGD.
>    - FLINK-1730 - Add a FlinkTools.persist style method to the Data Set.
>    https://issues.apache.org/jira/browse/FLINK-1730
>    This is again relevant to many algorithms, to create intermediate
>    results etc, for example L-BFGS development has been attempted 2-3
> times,
>    but always abandoned because of the need to collect a DataSet kills the
>    performance.
>    - FLINK-5782 - Support GPU calculations
>    https://issues.apache.org/jira/browse/FLINK-5782
>    Many algorithms will benefit greatly by GPU-accelerated linear algebra,
>    to the point where if a library doesn't support it puts it at a severe
>    disadvantage compared to other offerings.
>
>
> These issues aside, Stephan has mentioned recently the possibility of
> re-structuring the Flink project to allow for more flexibility for the
> libraries. I think that sounds quite promising and it should allow the
> development to pick up in the libraries, if we can get some more people
> reviewing and merging PRs.
>
> I would be all for updating our vision and roadmap to match what the
> community desires from the library.
>
> [1]
>
> https://cwiki.apache.org/confluence/display/FLINK/FlinkML%3A+Vision+and+Roadmap
>
> On Mon, Feb 20, 2017 at 12:47 PM, Gábor Hermann <[hidden email]>
> wrote:
>
> > Hi Stavros,
> >
> > Thanks for bringing this up.
> >
> > There have been past [1] and recent [2, 3] discussions about the Flink
> > libraries, because there are some stalling PRs and overloaded committers.
> > (Actually, Till is the only committer shepherd of the both the CEP and ML
> > library, and AFAIK he has a ton of other responsibilities and work to
> do.)
> > Thus it's hard to get code reviewed and merged, and without merged code
> > it's hard to get a committer status, so there are not many committers who
> > can review e.g. ML algorithm implementations, and the cycle goes on.
> Until
> > this is resolved somehow, we should help the committers by reviewing
> > each-others PRs.
> >
> > I think prioritizing features (b) is a good way to start. We could
> declare
> > most blocking features and concentrate on reviewing and merging them
> before
> > moving forward. E.g. the evaluation framework is quite important for an
> ML
> > library in my opinion, and has a PR stalling for long [4].
> >
> > Regarding c),  there are styleguides generally for contributing to Flink,
> > so we should follow that. Is there something more ML specific you think
> we
> > could follow? We should definitely declare, we follow scikit-learn and
> make
> > sure contributions comply to that.
> >
> > In terms of features (a, d), I think we should first see the bigger
> > picture. That is, it would be nice to discuss a clearer direction for
> Flink
> > ML. I've seen a lot of interest in contributing to Flink ML lately. I
> > believe we should rethink our goals, to put the contribution efforts in
> > making a usable and useful library. Are we trying to implement as many
> > useful algorithms as possible to create a scalable ML library? That would
> > seem ambitious, and of course there are a lot of frameworks and libraries
> > that already has something like this as goal (e.g. Spark MLlib, Mahout).
> > Should we rather create connectors to existing libraries? Then we cannot
> > really do Flink specific optimizations. Should we go for online machine
> > learning (as Flink is concentrating on streaming)? We already have a
> > connector to SAMOA. We could go on with questions like this. Maybe I'm
> > missing something, but I haven't seen such directions declared.
> >
> > Cheers,
> > Gabor
> >
> > [1] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > com/Opening-a-discussion-on-FlinkML-td10265.html
> > [2] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > com/Flink-CEP-development-is-stalling-td15237.html#a15341
> > [3] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > com/New-Flink-team-member-Kate-Eri-td15349.html
> > [4] https://github.com/apache/flink/pull/1849
> >
> >
> > On 2017-02-20 11:43, Stavros Kontopoulos wrote:
> >
> > (Resending with the appropriate topic)
> >>
> >> Hi,
> >>
> >> I would like to start a discussion about next steps for Flink ML.
> >> Currently there is a lot of work going on but needs a push forward.
> >>
> >> Some topics to discuss:
> >>
> >> a) How several features should be planned and get aligned with Flink
> >> releases.
> >> b) Priorities of what should be done.
> >> c) Basic guidelines for code: styleguides, scikit-learn compliance etc
> >> d) Missing features important for the success of the library, next steps
> >> etc...
> >>
> >> Thoughts?
> >>
> >> Best,
> >> Stavros
> >>
> >>
> >
>

[Timur Shenkao]

Hello guys,

My couple of cents.
All Flink presentations, articles, etc. articulate that Flink is for ETL,
data ingestion. CEP is a maximum.
If you visit http://flink.apache.org/usecases.html, you'll there aren't any
explicit ML or Graphs there.
It's also stated that Flink is suitable when "Data that is processed
quickly".
That's why people believe that Flink isn't for ML or don't even know that
Flink has such algorithms.
Then, folks decide: "I would better use old good Spark or scikit-learn than
dive into Flink's internals & implement algo by myself "

Sincerely yours, Timur

On Mon, Feb 20, 2017 at 1:53 PM, Katherin Eri <[hidden email]>
wrote:

> Hello guys,
>
>
> May be we will be able to focus our forces on some E2E scenario or show
> case for Flink as also ML supporting engine, and in such a way actualize
> the roadmap?
>
>
> This means: we can take some real life/production problem, like Fraud
> detection in some area, and try to solve this problem from the point of
> view of DataScience.
>
> Starting from data preprocessing and preparation, finishing
> implementation/usage of some ML algorithm.
>
> Doing this we will understand which issues are showstopper for
> implementation of such functionality. We will be able to understand Flink’s
> users better.
>
>
> May be community could share its ideas which show case could be the most
> useful for Apache Flink, or may be Data artisans could lead this?
>
> пн, 20 февр. 2017 г. в 15:28, Theodore Vasiloudis <
> [hidden email]>:
>
> > Hello all,
> >
> > thank you for opening this discussion Stavros, note that it's almost
> > exactly 1 year since I last opened such a topic (linked by Gabor) and the
> > comments there are still relevant.
> >
> > I think Gabor described the current state quite well, development in the
> > libraries is hard without committers dedicated to each project, and as a
> > result FlinkML and CEP have stalled.
> >
> > I think it's important to look at why development has stalled as well. As
> > people have mentioned there's a multitude of ML libraries out there and
> my
> > impression was that not many people are looking to use Flink for ML.
> Lately
> > that seems to have changed (with some interest shown in the Flink survey
> as
> > well).
> >
> > Gabor makes some good points about future directions for the library. Our
> > initial goal [1] was to make a truly scalable, easy to use library,
> within
> > the Flink ecosystem, providing a set of "workhorse" algorithms, sampled
> > from what's actually being used in the industry. We planned for a library
> > that has few algorithms, but does them properly.
> >
> > If we decide to go the way of focusing within Flink we face some major
> > challenges, because these are system limitations that do not necessarily
> > align with the goals of the community. Some issues relevant to ML on
> Flink
> > are:
> >
> >    - FLINK-2396 - Review the datasets of dynamic path and static path in
> >    iteration.
> >    https://issues.apache.org/jira/browse/FLINK-2396
> >    This has to do with the ability to iterate over one datset (model)
> while
> >    changing another (dataset), which is necessary for many ML algorithms
> > like
> >    SGD.
> >    - FLINK-1730 - Add a FlinkTools.persist style method to the Data Set.
> >    https://issues.apache.org/jira/browse/FLINK-1730
> >    This is again relevant to many algorithms, to create intermediate
> >    results etc, for example L-BFGS development has been attempted 2-3
> > times,
> >    but always abandoned because of the need to collect a DataSet kills
> the
> >    performance.
> >    - FLINK-5782 - Support GPU calculations
> >    https://issues.apache.org/jira/browse/FLINK-5782
> >    Many algorithms will benefit greatly by GPU-accelerated linear
> algebra,
> >    to the point where if a library doesn't support it puts it at a severe
> >    disadvantage compared to other offerings.
> >
> >
> > These issues aside, Stephan has mentioned recently the possibility of
> > re-structuring the Flink project to allow for more flexibility for the
> > libraries. I think that sounds quite promising and it should allow the
> > development to pick up in the libraries, if we can get some more people
> > reviewing and merging PRs.
> >
> > I would be all for updating our vision and roadmap to match what the
> > community desires from the library.
> >
> > [1]
> >
> > https://cwiki.apache.org/confluence/display/FLINK/
> FlinkML%3A+Vision+and+Roadmap
> >
> > On Mon, Feb 20, 2017 at 12:47 PM, Gábor Hermann <[hidden email]>
> > wrote:
> >
> > > Hi Stavros,
> > >
> > > Thanks for bringing this up.
> > >
> > > There have been past [1] and recent [2, 3] discussions about the Flink
> > > libraries, because there are some stalling PRs and overloaded
> committers.
> > > (Actually, Till is the only committer shepherd of the both the CEP and
> ML
> > > library, and AFAIK he has a ton of other responsibilities and work to
> > do.)
> > > Thus it's hard to get code reviewed and merged, and without merged code
> > > it's hard to get a committer status, so there are not many committers
> who
> > > can review e.g. ML algorithm implementations, and the cycle goes on.
> > Until
> > > this is resolved somehow, we should help the committers by reviewing
> > > each-others PRs.
> > >
> > > I think prioritizing features (b) is a good way to start. We could
> > declare
> > > most blocking features and concentrate on reviewing and merging them
> > before
> > > moving forward. E.g. the evaluation framework is quite important for an
> > ML
> > > library in my opinion, and has a PR stalling for long [4].
> > >
> > > Regarding c),  there are styleguides generally for contributing to
> Flink,
> > > so we should follow that. Is there something more ML specific you think
> > we
> > > could follow? We should definitely declare, we follow scikit-learn and
> > make
> > > sure contributions comply to that.
> > >
> > > In terms of features (a, d), I think we should first see the bigger
> > > picture. That is, it would be nice to discuss a clearer direction for
> > Flink
> > > ML. I've seen a lot of interest in contributing to Flink ML lately. I
> > > believe we should rethink our goals, to put the contribution efforts in
> > > making a usable and useful library. Are we trying to implement as many
> > > useful algorithms as possible to create a scalable ML library? That
> would
> > > seem ambitious, and of course there are a lot of frameworks and
> libraries
> > > that already has something like this as goal (e.g. Spark MLlib,
> Mahout).
> > > Should we rather create connectors to existing libraries? Then we
> cannot
> > > really do Flink specific optimizations. Should we go for online machine
> > > learning (as Flink is concentrating on streaming)? We already have a
> > > connector to SAMOA. We could go on with questions like this. Maybe I'm
> > > missing something, but I haven't seen such directions declared.
> > >
> > > Cheers,
> > > Gabor
> > >
> > > [1] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > com/Opening-a-discussion-on-FlinkML-td10265.html
> > > [2] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > com/Flink-CEP-development-is-stalling-td15237.html#a15341
> > > [3] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > com/New-Flink-team-member-Kate-Eri-td15349.html
> > > [4] https://github.com/apache/flink/pull/1849
> > >
> > >
> > > On 2017-02-20 11:43, Stavros Kontopoulos wrote:
> > >
> > > (Resending with the appropriate topic)
> > >>
> > >> Hi,
> > >>
> > >> I would like to start a discussion about next steps for Flink ML.
> > >> Currently there is a lot of work going on but needs a push forward.
> > >>
> > >> Some topics to discuss:
> > >>
> > >> a) How several features should be planned and get aligned with Flink
> > >> releases.
> > >> b) Priorities of what should be done.
> > >> c) Basic guidelines for code: styleguides, scikit-learn compliance etc
> > >> d) Missing features important for the success of the library, next
> steps
> > >> etc...
> > >>
> > >> Thoughts?
> > >>
> > >> Best,
> > >> Stavros
> > >>
> > >>
> > >
> >
>

[Stavros Kontopoulos]

I think Flink ML could be a success. Many use cases out there could benefit
from such algorithms especially online ones.
I agree examples should be created showing how it could be used.

I was not aware of the project re-structuring issues. GPUs is really
important nowdays but it is still not the major reason for not adopting
Flink ML. Flink ML has to be developed further and promoted as previously
stated.

In the meantime as for the reviewing part I am investing time there, so I
would like to see if we can join forces and push stuff.

I am aware of the evaluation framework PR and I will review it this week
hopefully. Bu can we commit on pushing anything given the load people have?

As another option could we propose someone to be the committer there as
well, someone Till will guide if it is needed?

I think we dont need to wait for all issues to be solved first. As for the
big picture re-use makes sense but I think the end result should be
something that benefits
Flink.  I would like to stay in Flink as much as possible from a
UX/features side of view. Of course people already use a number of
libraries for years and what we do by implementing the algorithms is
getting those algorithms to work on large datasets plus for streaming,
keeping the UX familiar at the same time.

I think connecting to external libraries should be done if possible for
things not being your domain like dbs or dfs etc... Is it a domain related
for a streaming engine? Use cases drive that IMHO... Again implementation
should be justified by user needs, if there is no such need no reason to
implement anything.

Just some thoughts...


On Mon, Feb 20, 2017 at 3:39 PM, Timur Shenkao <[hidden email]> wrote:

> Hello guys,
>
> My couple of cents.
> All Flink presentations, articles, etc. articulate that Flink is for ETL,
> data ingestion. CEP is a maximum.
> If you visit http://flink.apache.org/usecases.html, you'll there aren't
> any
> explicit ML or Graphs there.
> It's also stated that Flink is suitable when "Data that is processed
> quickly".
> That's why people believe that Flink isn't for ML or don't even know that
> Flink has such algorithms.
> Then, folks decide: "I would better use old good Spark or scikit-learn than
> dive into Flink's internals & implement algo by myself "
>
> Sincerely yours, Timur
>
> On Mon, Feb 20, 2017 at 1:53 PM, Katherin Eri <[hidden email]>
> wrote:
>
> > Hello guys,
> >
> >
> > May be we will be able to focus our forces on some E2E scenario or show
> > case for Flink as also ML supporting engine, and in such a way actualize
> > the roadmap?
> >
> >
> > This means: we can take some real life/production problem, like Fraud
> > detection in some area, and try to solve this problem from the point of
> > view of DataScience.
> >
> > Starting from data preprocessing and preparation, finishing
> > implementation/usage of some ML algorithm.
> >
> > Doing this we will understand which issues are showstopper for
> > implementation of such functionality. We will be able to understand
> Flink’s
> > users better.
> >
> >
> > May be community could share its ideas which show case could be the most
> > useful for Apache Flink, or may be Data artisans could lead this?
> >
> > пн, 20 февр. 2017 г. в 15:28, Theodore Vasiloudis <
> > [hidden email]>:
> >
> > > Hello all,
> > >
> > > thank you for opening this discussion Stavros, note that it's almost
> > > exactly 1 year since I last opened such a topic (linked by Gabor) and
> the
> > > comments there are still relevant.
> > >
> > > I think Gabor described the current state quite well, development in
> the
> > > libraries is hard without committers dedicated to each project, and as
> a
> > > result FlinkML and CEP have stalled.
> > >
> > > I think it's important to look at why development has stalled as well.
> As
> > > people have mentioned there's a multitude of ML libraries out there and
> > my
> > > impression was that not many people are looking to use Flink for ML.
> > Lately
> > > that seems to have changed (with some interest shown in the Flink
> survey
> > as
> > > well).
> > >
> > > Gabor makes some good points about future directions for the library.
> Our
> > > initial goal [1] was to make a truly scalable, easy to use library,
> > within
> > > the Flink ecosystem, providing a set of "workhorse" algorithms, sampled
> > > from what's actually being used in the industry. We planned for a
> library
> > > that has few algorithms, but does them properly.
> > >
> > > If we decide to go the way of focusing within Flink we face some major
> > > challenges, because these are system limitations that do not
> necessarily
> > > align with the goals of the community. Some issues relevant to ML on
> > Flink
> > > are:
> > >
> > >    - FLINK-2396 - Review the datasets of dynamic path and static path
> in
> > >    iteration.
> > >    https://issues.apache.org/jira/browse/FLINK-2396
> > >    This has to do with the ability to iterate over one datset (model)
> > while
> > >    changing another (dataset), which is necessary for many ML
> algorithms
> > > like
> > >    SGD.
> > >    - FLINK-1730 - Add a FlinkTools.persist style method to the Data
> Set.
> > >    https://issues.apache.org/jira/browse/FLINK-1730
> > >    This is again relevant to many algorithms, to create intermediate
> > >    results etc, for example L-BFGS development has been attempted 2-3
> > > times,
> > >    but always abandoned because of the need to collect a DataSet kills
> > the
> > >    performance.
> > >    - FLINK-5782 - Support GPU calculations
> > >    https://issues.apache.org/jira/browse/FLINK-5782
> > >    Many algorithms will benefit greatly by GPU-accelerated linear
> > algebra,
> > >    to the point where if a library doesn't support it puts it at a
> severe
> > >    disadvantage compared to other offerings.
> > >
> > >
> > > These issues aside, Stephan has mentioned recently the possibility of
> > > re-structuring the Flink project to allow for more flexibility for the
> > > libraries. I think that sounds quite promising and it should allow the
> > > development to pick up in the libraries, if we can get some more people
> > > reviewing and merging PRs.
> > >
> > > I would be all for updating our vision and roadmap to match what the
> > > community desires from the library.
> > >
> > > [1]
> > >
> > > https://cwiki.apache.org/confluence/display/FLINK/
> > FlinkML%3A+Vision+and+Roadmap
> > >
> > > On Mon, Feb 20, 2017 at 12:47 PM, Gábor Hermann <[hidden email]
> >
> > > wrote:
> > >
> > > > Hi Stavros,
> > > >
> > > > Thanks for bringing this up.
> > > >
> > > > There have been past [1] and recent [2, 3] discussions about the
> Flink
> > > > libraries, because there are some stalling PRs and overloaded
> > committers.
> > > > (Actually, Till is the only committer shepherd of the both the CEP
> and
> > ML
> > > > library, and AFAIK he has a ton of other responsibilities and work to
> > > do.)
> > > > Thus it's hard to get code reviewed and merged, and without merged
> code
> > > > it's hard to get a committer status, so there are not many committers
> > who
> > > > can review e.g. ML algorithm implementations, and the cycle goes on.
> > > Until
> > > > this is resolved somehow, we should help the committers by reviewing
> > > > each-others PRs.
> > > >
> > > > I think prioritizing features (b) is a good way to start. We could
> > > declare
> > > > most blocking features and concentrate on reviewing and merging them
> > > before
> > > > moving forward. E.g. the evaluation framework is quite important for
> an
> > > ML
> > > > library in my opinion, and has a PR stalling for long [4].
> > > >
> > > > Regarding c),  there are styleguides generally for contributing to
> > Flink,
> > > > so we should follow that. Is there something more ML specific you
> think
> > > we
> > > > could follow? We should definitely declare, we follow scikit-learn
> and
> > > make
> > > > sure contributions comply to that.
> > > >
> > > > In terms of features (a, d), I think we should first see the bigger
> > > > picture. That is, it would be nice to discuss a clearer direction for
> > > Flink
> > > > ML. I've seen a lot of interest in contributing to Flink ML lately. I
> > > > believe we should rethink our goals, to put the contribution efforts
> in
> > > > making a usable and useful library. Are we trying to implement as
> many
> > > > useful algorithms as possible to create a scalable ML library? That
> > would
> > > > seem ambitious, and of course there are a lot of frameworks and
> > libraries
> > > > that already has something like this as goal (e.g. Spark MLlib,
> > Mahout).
> > > > Should we rather create connectors to existing libraries? Then we
> > cannot
> > > > really do Flink specific optimizations. Should we go for online
> machine
> > > > learning (as Flink is concentrating on streaming)? We already have a
> > > > connector to SAMOA. We could go on with questions like this. Maybe
> I'm
> > > > missing something, but I haven't seen such directions declared.
> > > >
> > > > Cheers,
> > > > Gabor
> > > >
> > > > [1] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > > com/Opening-a-discussion-on-FlinkML-td10265.html
> > > > [2] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > > com/Flink-CEP-development-is-stalling-td15237.html#a15341
> > > > [3] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > > com/New-Flink-team-member-Kate-Eri-td15349.html
> > > > [4] https://github.com/apache/flink/pull/1849
> > > >
> > > >
> > > > On 2017-02-20 11:43, Stavros Kontopoulos wrote:
> > > >
> > > > (Resending with the appropriate topic)
> > > >>
> > > >> Hi,
> > > >>
> > > >> I would like to start a discussion about next steps for Flink ML.
> > > >> Currently there is a lot of work going on but needs a push forward.
> > > >>
> > > >> Some topics to discuss:
> > > >>
> > > >> a) How several features should be planned and get aligned with Flink
> > > >> releases.
> > > >> b) Priorities of what should be done.
> > > >> c) Basic guidelines for code: styleguides, scikit-learn compliance
> etc
> > > >> d) Missing features important for the success of the library, next
> > steps
> > > >> etc...
> > > >>
> > > >> Thoughts?
> > > >>
> > > >> Best,
> > > >> Stavros
> > > >>
> > > >>
> > > >
> > >
> >
>

[Andrea Spina]

Hi all,

Thanks Stavros for pushing forward the discussion which I feel really relevant.

Since I'm approaching actively the community just right now and I haven't enough experience and such visibility around the Flink community, I'd limit myself to share an opinion as a Flink user.

I'm using Flink since almost a year along two different experiences, but I've bumped into the question "how to handle ML workloads and keep Flink as the main engine?" in both cases. Then the first point raises in my mind: why do I need to adopt an extra system for purely ML purposes: how amazing could be to benefit the Flink engine as ML features provider and to avoid paying the effort to maintain an additional engine? This thought links also @Timur opinion: I believe that users would prefer way more a unified architecture in this case. Even if a user want to use an external tool/library - perhaps providing additional language support (e.g. R) - so that user should be capable to run it on top of Flink.

Along my work with Flink I needed to implement some ML algorithms on both Flink and Spark and I often struggled with Flink performances: namely, I think (in the name of the bigger picture) we should first focus the effort on solving some well-known Flink limitations as @theodore pinpointed. I'd like to highlight [1] and [2] which I find relevant. Since the community would decide to go ahead with FlinkML I believe fixing the above described issues may be a good starting point. That would also definitely push forward some important integrations as Apache SystemML.

Given all these points, I'm increasingly convinced that Online Machine Learning would be the real final objective and the more suitable goal since we're talking about a real-time streaming engine and - from a real high point of view - I believe Flink would fit this topic in a more genuine way than the batch case. We've a connector for Apache SAMOA, but it seems in an early stage of development IMHO and not really active. If we want to make something within Flink instead, we need to speed up the design of some features (e.g. side inputs [3]).

I really hope we can define a new roadmap by which we can finally push forward the topic. I will put my best to help in this way.

Sincerely,
Andrea

[1] Add a FlinkTools.persist style method to the Data Set
https://issues.apache.org/jira/browse/FLINK-1730
[2] Only send data to each taskmanager once for broadcasts
https://cwiki.apache.org/confluence/display/FLINK/FLIP-5%3A+Only+send+data+to+each+taskmanager+once+for+broadcasts
[3] Side inputs - Evolving or static Filter/Enriching
https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-MKQYN3m4/edit#
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Add-Side-Input-Broadcast-Set-For-Streaming-API-td11529.html

[Theodore Vasiloudis]

Thank you all for your thoughts on the matter.

Andrea brought up some further engine considerations that we need to
address in order to have a competitive ML engine on Flink.

I'm happy to see many people willing to contribute to the development of ML
on Flink. The way I see it, there needs to be buy-in from the rest of the
community for such changes to go through.

If then you are interested in helping out, tackling one of the issues
mentioned in my previous email or the ones mentioned by Andrea are the most
critical ones, as they require making changes to the core.

If you want to take up one of those issues the best way is to start a
conversation on the list, and gauge the opinion of the community.

Finally, as Stavros mentioned, we need to come up with an updated roadmap
for FlinkML that includes these issues.

@Andrea, the idea of an online learning library for Flink has been broached
before, and this semester I have one Master student working on exactly
that. From my conversations with people in the industry however, almost
nobody uses online learning in production, at best models are updated every
5 minutes. So the impact would probably not be very large.

I would like to bring up again the topic of model serving that I think fits
the Flink use-case much better. Developing a system like Clipper [1] on top
of Flink could be one of the best ways to use Flink for ML.

Regards,
Theodore

[1]  Clipper: A Low-Latency Online Prediction Serving System -
https://arxiv.org/abs/1612.03079

On Tue, Feb 21, 2017 at 12:10 AM, Andrea Spina <[hidden email]>
wrote:

> Hi all,
>
> Thanks Stavros for pushing forward the discussion which I feel really
> relevant.
>
> Since I'm approaching actively the community just right now and I haven't
> enough experience and such visibility around the Flink community, I'd limit
> myself to share an opinion as a Flink user.
>
> I'm using Flink since almost a year along two different experiences, but
> I've bumped into the question "how to handle ML workloads and keep Flink as
> the main engine?" in both cases. Then the first point raises in my mind:
> why
> do I need to adopt an extra system for purely ML purposes: how amazing
> could
> be to benefit the Flink engine as ML features provider and to avoid paying
> the effort to maintain an additional engine? This thought links also @Timur
> opinion: I believe that users would prefer way more a unified architecture
> in this case. Even if a user want to use an external tool/library - perhaps
> providing additional language support (e.g. R) - so that user should be
> capable to run it on top of Flink.
>
> Along my work with Flink I needed to implement some ML algorithms on both
> Flink and Spark and I often struggled with Flink performances: namely, I
> think (in the name of the bigger picture) we should first focus the effort
> on solving some well-known Flink limitations as @theodore pinpointed. I'd
> like to highlight [1] and [2] which I find relevant. Since the community
> would decide to go ahead with FlinkML I believe fixing the above described
> issues may be a good starting point. That would also definitely push
> forward
> some important integrations as Apache SystemML.
>
> Given all these points, I'm increasingly convinced that Online Machine
> Learning would be the real final objective and the more suitable goal since
> we're talking about a real-time streaming engine and - from a real high
> point of view - I believe Flink would fit this topic in a more genuine way
> than the batch case. We've a connector for Apache SAMOA, but it seems in an
> early stage of development IMHO and not really active. If we want to make
> something within Flink instead, we need to speed up the design of some
> features (e.g. side inputs [3]).
>
> I really hope we can define a new roadmap by which we can finally push
> forward the topic. I will put my best to help in this way.
>
> Sincerely,
> Andrea
>
> [1] Add a FlinkTools.persist style method to the Data Set
> https://issues.apache.org/jira/browse/FLINK-1730
> [2] Only send data to each taskmanager once for broadcasts
> https://cwiki.apache.org/confluence/display/FLINK/FLIP-
> 5%3A+Only+send+data+to+each+taskmanager+once+for+broadcasts
> [3] Side inputs - Evolving or static Filter/Enriching
> https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-
> MKQYN3m4/edit#
> http://apache-flink-mailing-list-archive.1008284.n3.
> nabble.com/DISCUSS-Add-Side-Input-Broadcast-Set-For-
> Streaming-API-td11529.html
>
>
>
> --
> View this message in context: http://apache-flink-mailing-
> list-archive.1008284.n3.nabble.com/DISCUSS-Flink-ML-
> roadmap-tp16040p16064.html
> Sent from the Apache Flink Mailing List archive. mailing list archive at
> Nabble.com.
>

[Till Rohrmann]

Thanks a lot for all your valuable input. It's great to see all your
interest in Flink and its ML library :-)

1) Direction of FlinkML

In order to reboot the FlinkML library we should indeed first decide on its
direction and come up with a roadmap to get the community behind.

Since we only have limited resources the question for me is first of all
whether we continue developing a batch ML library or whether we concentrate
on streaming machine learning.

The core idea of FlinkML was to provide the user with an easy toolbox to
create machine learning pipelines. These pipelines are per se not batch or
streaming specific but so far all our implementations are based on Flink's
batch API.

While implementing the ML algorithms we realized that Flink's engine has
still some deficiencies on the batch side. Theo already mentioned the
iteration problem with static inputs [1] and the problem of caching
intermediate results [2]. But there are also other problems such as dynamic
memory management [3] and a leg wise scheduling [4] for complex topologies.
Without these features, I don't see that Flink will be able to efficiently
execute batch ML jobs. Unfortunately, all of these problems are far from
trivial to solve and will require quite some changes to Flink's runtime.
Given Flink's current focus on stream processing, I don't see enough
community capacities left to implement these features soon.

Furthermore, if we decide to continue pursuing the batch direction, then
we'll be in direct competition with more established frameworks such as
SparkML, Weka, TensorFlow and scikit-learn, for example. I guess that alone
the work to catch up with these libraries in terms of algorithm support
will be quite challenging.

Therefore, I think it would be more promising to concentrate on streaming
ML and try to establish Flink's brand there. Streaming ML has not been as
thoroughly explored as the batch counterpart and there are not too many
players on the field. Furthermore, it would be well aligned with the
direction of the rest of the project.

1.1) Possible features

I agree with Theo that model serving/low latency prediction would be a
really good/almost natural use case for Flink. For that we would need to be
able to import trained models and do predictions with them. Maybe Clipper
is a good solution for that or maybe PMML or another model format. That is
something we would have to research.

Next, in order to support continuous model updates (maybe from a
periodically triggered batch job) we would need side input support.

With these two features we could probably already realize some really cool
use cases.

2) Growing Flink's ML community

One of the problems with FlinkML, as you've mentioned it, was the lack of
active committer support after the initial development. As Gabor pointed
out if there is no committer around then there is only little chance to
become one if nothing gets merged, even though we're in heavy need for them.

Since I'm the culprit in this case, I can tell you that it would be
tremendously helpful if the community (including in our case mostly
contributors) continues reviewing actively each others PRs. If a PR is in
good shape than it's much easier (less work) for to merge it. I think this
could be an immediate action point.

Next, I started a discussion thread [5] about restructuring Flink in order
to decrease test and build times but also to allow adding new committers
more easily for modules where we have a high need. Maybe this can help to
solve the committer problem.

3) Showcasing capabilities

I agree with Timur's observation that we have far too little material out
there which showcases what's actually possible to do with Flink wrt ML.
That is something which we can start right away to change. One good
possibility is always to write a blog post about an interesting use case
you've implemented. Thus, I like very much Katherin's idea. And indeed when
I implemented the ALS matrix factorization with Flink, we came across a lot
of problems with Flink.

The other good option which was mentioned is the creation of a kind of ML
cookbook. The cookbook could contain advanced recipes how to solve certain
problems with FlinkML. The Flink community always wanted to create such a
cookbook for Flink in general. Maybe we could lay the first foundation for
it.

[1] https://issues.apache.org/jira/browse/FLINK-2396
[2] https://issues.apache.org/jira/browse/FLINK-1404
[3] https://issues.apache.org/jira/browse/FLINK-1101
[4] https://issues.apache.org/jira/browse/FLINK-2119
[5]
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Project-build-time-and-possible-restructuring-tt16088.html

Cheers,
Till

On Tue, Feb 21, 2017 at 12:04 PM, Theodore Vasiloudis <
[hidden email]> wrote:

> Thank you all for your thoughts on the matter.
>
> Andrea brought up some further engine considerations that we need to
> address in order to have a competitive ML engine on Flink.
>
> I'm happy to see many people willing to contribute to the development of ML
> on Flink. The way I see it, there needs to be buy-in from the rest of the
> community for such changes to go through.
>
> If then you are interested in helping out, tackling one of the issues
> mentioned in my previous email or the ones mentioned by Andrea are the most
> critical ones, as they require making changes to the core.
>
> If you want to take up one of those issues the best way is to start a
> conversation on the list, and gauge the opinion of the community.
>
> Finally, as Stavros mentioned, we need to come up with an updated roadmap
> for FlinkML that includes these issues.
>
> @Andrea, the idea of an online learning library for Flink has been broached
> before, and this semester I have one Master student working on exactly
> that. From my conversations with people in the industry however, almost
> nobody uses online learning in production, at best models are updated every
> 5 minutes. So the impact would probably not be very large.
>
> I would like to bring up again the topic of model serving that I think fits
> the Flink use-case much better. Developing a system like Clipper [1] on top
> of Flink could be one of the best ways to use Flink for ML.
>
> Regards,
> Theodore
>
> [1]  Clipper: A Low-Latency Online Prediction Serving System -
> https://arxiv.org/abs/1612.03079
>
> On Tue, Feb 21, 2017 at 12:10 AM, Andrea Spina <[hidden email]
> >
> wrote:
>
> > Hi all,
> >
> > Thanks Stavros for pushing forward the discussion which I feel really
> > relevant.
> >
> > Since I'm approaching actively the community just right now and I haven't
> > enough experience and such visibility around the Flink community, I'd
> limit
> > myself to share an opinion as a Flink user.
> >
> > I'm using Flink since almost a year along two different experiences, but
> > I've bumped into the question "how to handle ML workloads and keep Flink
> as
> > the main engine?" in both cases. Then the first point raises in my mind:
> > why
> > do I need to adopt an extra system for purely ML purposes: how amazing
> > could
> > be to benefit the Flink engine as ML features provider and to avoid
> paying
> > the effort to maintain an additional engine? This thought links also
> @Timur
> > opinion: I believe that users would prefer way more a unified
> architecture
> > in this case. Even if a user want to use an external tool/library -
> perhaps
> > providing additional language support (e.g. R) - so that user should be
> > capable to run it on top of Flink.
> >
> > Along my work with Flink I needed to implement some ML algorithms on both
> > Flink and Spark and I often struggled with Flink performances: namely, I
> > think (in the name of the bigger picture) we should first focus the
> effort
> > on solving some well-known Flink limitations as @theodore pinpointed. I'd
> > like to highlight [1] and [2] which I find relevant. Since the community
> > would decide to go ahead with FlinkML I believe fixing the above
> described
> > issues may be a good starting point. That would also definitely push
> > forward
> > some important integrations as Apache SystemML.
> >
> > Given all these points, I'm increasingly convinced that Online Machine
> > Learning would be the real final objective and the more suitable goal
> since
> > we're talking about a real-time streaming engine and - from a real high
> > point of view - I believe Flink would fit this topic in a more genuine
> way
> > than the batch case. We've a connector for Apache SAMOA, but it seems in
> an
> > early stage of development IMHO and not really active. If we want to make
> > something within Flink instead, we need to speed up the design of some
> > features (e.g. side inputs [3]).
> >
> > I really hope we can define a new roadmap by which we can finally push
> > forward the topic. I will put my best to help in this way.
> >
> > Sincerely,
> > Andrea
> >
> > [1] Add a FlinkTools.persist style method to the Data Set
> > https://issues.apache.org/jira/browse/FLINK-1730
> > [2] Only send data to each taskmanager once for broadcasts
> > https://cwiki.apache.org/confluence/display/FLINK/FLIP-
> > 5%3A+Only+send+data+to+each+taskmanager+once+for+broadcasts
> > [3] Side inputs - Evolving or static Filter/Enriching
> > https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-
> > MKQYN3m4/edit#
> > http://apache-flink-mailing-list-archive.1008284.n3.
> > nabble.com/DISCUSS-Add-Side-Input-Broadcast-Set-For-
> > Streaming-API-td11529.html
> >
> >
> >
> > --
> > View this message in context: http://apache-flink-mailing-
> > list-archive.1008284.n3.nabble.com/DISCUSS-Flink-ML-
> > roadmap-tp16040p16064.html
> > Sent from the Apache Flink Mailing List archive. mailing list archive at
> > Nabble.com.
> >
>

[Stavros Kontopoulos]

Ok I see. Suppose we solve all the critical issues. And suppose we dont go
with the pure online model (although online ML has a potential)... should
we move on with the
current ML implementation which is for batch processing (to the best of my
knowledge)? The parameter server problem is a long standing one and many
companies out there started to provide their own solutions. That would be
very useful but I see it only as part of the solution.

The other thing is that when someone is working locally and does some work
with Flink he should need to go out of it to play with other libraries.
Isnt this important for the product success?

Regards,
Stavros
On Tue, Feb 21, 2017 at 1:04 PM, Theodore Vasiloudis <
[hidden email]> wrote:

> Thank you all for your thoughts on the matter.
>
> Andrea brought up some further engine considerations that we need to
> address in order to have a competitive ML engine on Flink.
>
> I'm happy to see many people willing to contribute to the development of ML
> on Flink. The way I see it, there needs to be buy-in from the rest of the
> community for such changes to go through.
>
> If then you are interested in helping out, tackling one of the issues
> mentioned in my previous email or the ones mentioned by Andrea are the most
> critical ones, as they require making changes to the core.
>
> If you want to take up one of those issues the best way is to start a
> conversation on the list, and gauge the opinion of the community.
>
> Finally, as Stavros mentioned, we need to come up with an updated roadmap
> for FlinkML that includes these issues.
>
> @Andrea, the idea of an online learning library for Flink has been broached
> before, and this semester I have one Master student working on exactly
> that. From my conversations with people in the industry however, almost
> nobody uses online learning in production, at best models are updated every
> 5 minutes. So the impact would probably not be very large.
>
> I would like to bring up again the topic of model serving that I think fits
> the Flink use-case much better. Developing a system like Clipper [1] on top
> of Flink could be one of the best ways to use Flink for ML.
>
> Regards,
> Theodore
>
> [1]  Clipper: A Low-Latency Online Prediction Serving System -
> https://arxiv.org/abs/1612.03079
>
> On Tue, Feb 21, 2017 at 12:10 AM, Andrea Spina <[hidden email]
> >
> wrote:
>
> > Hi all,
> >
> > Thanks Stavros for pushing forward the discussion which I feel really
> > relevant.
> >
> > Since I'm approaching actively the community just right now and I haven't
> > enough experience and such visibility around the Flink community, I'd
> limit
> > myself to share an opinion as a Flink user.
> >
> > I'm using Flink since almost a year along two different experiences, but
> > I've bumped into the question "how to handle ML workloads and keep Flink
> as
> > the main engine?" in both cases. Then the first point raises in my mind:
> > why
> > do I need to adopt an extra system for purely ML purposes: how amazing
> > could
> > be to benefit the Flink engine as ML features provider and to avoid
> paying
> > the effort to maintain an additional engine? This thought links also
> @Timur
> > opinion: I believe that users would prefer way more a unified
> architecture
> > in this case. Even if a user want to use an external tool/library -
> perhaps
> > providing additional language support (e.g. R) - so that user should be
> > capable to run it on top of Flink.
> >
> > Along my work with Flink I needed to implement some ML algorithms on both
> > Flink and Spark and I often struggled with Flink performances: namely, I
> > think (in the name of the bigger picture) we should first focus the
> effort
> > on solving some well-known Flink limitations as @theodore pinpointed. I'd
> > like to highlight [1] and [2] which I find relevant. Since the community
> > would decide to go ahead with FlinkML I believe fixing the above
> described
> > issues may be a good starting point. That would also definitely push
> > forward
> > some important integrations as Apache SystemML.
> >
> > Given all these points, I'm increasingly convinced that Online Machine
> > Learning would be the real final objective and the more suitable goal
> since
> > we're talking about a real-time streaming engine and - from a real high
> > point of view - I believe Flink would fit this topic in a more genuine
> way
> > than the batch case. We've a connector for Apache SAMOA, but it seems in
> an
> > early stage of development IMHO and not really active. If we want to make
> > something within Flink instead, we need to speed up the design of some
> > features (e.g. side inputs [3]).
> >
> > I really hope we can define a new roadmap by which we can finally push
> > forward the topic. I will put my best to help in this way.
> >
> > Sincerely,
> > Andrea
> >
> > [1] Add a FlinkTools.persist style method to the Data Set
> > https://issues.apache.org/jira/browse/FLINK-1730
> > [2] Only send data to each taskmanager once for broadcasts
> > https://cwiki.apache.org/confluence/display/FLINK/FLIP-
> > 5%3A+Only+send+data+to+each+taskmanager+once+for+broadcasts
> > [3] Side inputs - Evolving or static Filter/Enriching
> > https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-
> > MKQYN3m4/edit#
> > http://apache-flink-mailing-list-archive.1008284.n3.
> > nabble.com/DISCUSS-Add-Side-Input-Broadcast-Set-For-
> > Streaming-API-td11529.html
> >
> >
> >
> > --
> > View this message in context: http://apache-flink-mailing-
> > list-archive.1008284.n3.nabble.com/DISCUSS-Flink-ML-
> > roadmap-tp16040p16064.html
> > Sent from the Apache Flink Mailing List archive. mailing list archive at
> > Nabble.com.
> >
>

[Katherin Eri]

Till, thank you for your response.
But I need several points to clarify:

1) Yes, batch and batch ML is the field full of alternatives, but in my
opinion that doesn’t mean that we should ignore the problem of not
developing batch part of Flink. You know: Apache Beam, Apache Mahout they
both feel the lack of normally implemented batching feature. DL4J will be
able to integrate with Apache Flink, but this integration will work only on
paper, and not efficient in production.

Did you mean with this phrase: “*Unfortunately, all of these problems are
far from trivial to solve and will require quite some changes to Flink's
runtime. Given Flink's current focus on stream processing, I don't see
enough community capacities left to implement these features soon*.”, that
Apache Flink won’t pay attention to batch part of it, or I have got you
wrong?

2) Yes, reimplementing libraries that already were developed by community
is not a good way, but maybe we should make from Flink engine that could
easily run ML libraries on top of it: integrate with SystemML, DL4J and
many many others? But doing this we well still required batch calculations.

вт, 21 февр. 2017 г. в 18:01, Stavros Kontopoulos <[hidden email]
>:

> Ok I see. Suppose we solve all the critical issues. And suppose we dont go
> with the pure online model (although online ML has a potential)... should
> we move on with the
> current ML implementation which is for batch processing (to the best of my
> knowledge)? The parameter server problem is a long standing one and many
> companies out there started to provide their own solutions. That would be
> very useful but I see it only as part of the solution.
>
> The other thing is that when someone is working locally and does some work
> with Flink he should need to go out of it to play with other libraries.
> Isnt this important for the product success?
>
> Regards,
> Stavros
> On Tue, Feb 21, 2017 at 1:04 PM, Theodore Vasiloudis <
> [hidden email]> wrote:
>
> > Thank you all for your thoughts on the matter.
> >
> > Andrea brought up some further engine considerations that we need to
> > address in order to have a competitive ML engine on Flink.
> >
> > I'm happy to see many people willing to contribute to the development of
> ML
> > on Flink. The way I see it, there needs to be buy-in from the rest of the
> > community for such changes to go through.
> >
> > If then you are interested in helping out, tackling one of the issues
> > mentioned in my previous email or the ones mentioned by Andrea are the
> most
> > critical ones, as they require making changes to the core.
> >
> > If you want to take up one of those issues the best way is to start a
> > conversation on the list, and gauge the opinion of the community.
> >
> > Finally, as Stavros mentioned, we need to come up with an updated roadmap
> > for FlinkML that includes these issues.
> >
> > @Andrea, the idea of an online learning library for Flink has been
> broached
> > before, and this semester I have one Master student working on exactly
> > that. From my conversations with people in the industry however, almost
> > nobody uses online learning in production, at best models are updated
> every
> > 5 minutes. So the impact would probably not be very large.
> >
> > I would like to bring up again the topic of model serving that I think
> fits
> > the Flink use-case much better. Developing a system like Clipper [1] on
> top
> > of Flink could be one of the best ways to use Flink for ML.
> >
> > Regards,
> > Theodore
> >
> > [1]  Clipper: A Low-Latency Online Prediction Serving System -
> > https://arxiv.org/abs/1612.03079
> >
> > On Tue, Feb 21, 2017 at 12:10 AM, Andrea Spina <
> [hidden email]
> > >
> > wrote:
> >
> > > Hi all,
> > >
> > > Thanks Stavros for pushing forward the discussion which I feel really
> > > relevant.
> > >
> > > Since I'm approaching actively the community just right now and I
> haven't
> > > enough experience and such visibility around the Flink community, I'd
> > limit
> > > myself to share an opinion as a Flink user.
> > >
> > > I'm using Flink since almost a year along two different experiences,
> but
> > > I've bumped into the question "how to handle ML workloads and keep
> Flink
> > as
> > > the main engine?" in both cases. Then the first point raises in my
> mind:
> > > why
> > > do I need to adopt an extra system for purely ML purposes: how amazing
> > > could
> > > be to benefit the Flink engine as ML features provider and to avoid
> > paying
> > > the effort to maintain an additional engine? This thought links also
> > @Timur
> > > opinion: I believe that users would prefer way more a unified
> > architecture
> > > in this case. Even if a user want to use an external tool/library -
> > perhaps
> > > providing additional language support (e.g. R) - so that user should be
> > > capable to run it on top of Flink.
> > >
> > > Along my work with Flink I needed to implement some ML algorithms on
> both
> > > Flink and Spark and I often struggled with Flink performances: namely,
> I
> > > think (in the name of the bigger picture) we should first focus the
> > effort
> > > on solving some well-known Flink limitations as @theodore pinpointed.
> I'd
> > > like to highlight [1] and [2] which I find relevant. Since the
> community
> > > would decide to go ahead with FlinkML I believe fixing the above
> > described
> > > issues may be a good starting point. That would also definitely push
> > > forward
> > > some important integrations as Apache SystemML.
> > >
> > > Given all these points, I'm increasingly convinced that Online Machine
> > > Learning would be the real final objective and the more suitable goal
> > since
> > > we're talking about a real-time streaming engine and - from a real high
> > > point of view - I believe Flink would fit this topic in a more genuine
> > way
> > > than the batch case. We've a connector for Apache SAMOA, but it seems
> in
> > an
> > > early stage of development IMHO and not really active. If we want to
> make
> > > something within Flink instead, we need to speed up the design of some
> > > features (e.g. side inputs [3]).
> > >
> > > I really hope we can define a new roadmap by which we can finally push
> > > forward the topic. I will put my best to help in this way.
> > >
> > > Sincerely,
> > > Andrea
> > >
> > > [1] Add a FlinkTools.persist style method to the Data Set
> > > https://issues.apache.org/jira/browse/FLINK-1730
> > > [2] Only send data to each taskmanager once for broadcasts
> > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-
> > > 5%3A+Only+send+data+to+each+taskmanager+once+for+broadcasts
> > > [3] Side inputs - Evolving or static Filter/Enriching
> > >
> https://docs.google.com/document/d/1hIgxi2Zchww_5fWUHLoYiXwSBXjv-M5eOv-
> > > MKQYN3m4/edit#
> > > http://apache-flink-mailing-list-archive.1008284.n3.
> > > nabble.com/DISCUSS-Add-Side-Input-Broadcast-Set-For-
> > > Streaming-API-td11529.html
> > >
> > >
> > >
> > > --
> > > View this message in context: http://apache-flink-mailing-
> > > list-archive.1008284.n3.nabble.com/DISCUSS-Flink-ML-
> > > roadmap-tp16040p16064.html
> > > Sent from the Apache Flink Mailing List archive. mailing list archive
> at
> > > Nabble.com.
> > >
> >
>

--

*Yours faithfully, *

*Kate Eri.*

[Gábor Hermann]

It's great to see so much activity in this discussion :)
I'll try to add my thoughts.

I think building a developer community (Till's 2. point) can be slightly
separated from what features we should aim for (1. point) and showcasing
(3. point). Thanks Till for bringing up the ideas for restructuring, I'm
sure we'll find a way to make the development process more dynamic. I'll
try to address the rest here.

It's hard to choose directions between streaming and batch ML. As Theo
has indicated, not much online ML is used in production, but Flink
concentrates on streaming, so online ML would be a better fit for Flink.
However, as most of you argued, there's definite need for batch ML. But
batch ML seems hard to achieve because there are blocking issues with
persisting, iteration paths etc. So it's no good either way.

I propose a seemingly crazy solution: what if we developed batch
algorithms also with the streaming API? The batch API would clearly seem
more suitable for ML algorithms, but there a lot of benefits of this
approach too, so it's clearly worth considering. Flink also has the high
level vision of "streaming for everything" that would clearly fit this
case. What do you all think about this? Do you think this solution would
be feasible? I would be happy to make a more elaborate proposal, but I
push my main ideas here:

1) Simplifying by using one system
It could simplify the work of both the users and the developers. One
could execute training once, or could execute it periodically e.g. by
using windows. Low-latency serving and training could be done in the
same system. We could implement incremental algorithms, without any side
inputs for combining online learning (or predictions) with batch
learning. Of course, all the logic describing these must be somehow
implemented (e.g. synchronizing predictions with training), but it
should be easier to do so in one system, than by combining e.g. the
batch and streaming API.

2) Batch ML with the streaming API is not harder
Despite these benefits, it could seem harder to implement batch ML with
the streaming API, but in my opinion it's not. There are more flexible,
lower-level optimization potentials with the streaming API. Most
distributed ML algorithms use a lower-level model than the batch API
anyway, so sometimes it feels like forcing the algorithm logic into the
training API and tweaking it. Although we could not use the batch
primitives like join, we would have the E.g. in my experience with
implementing a distributed matrix factorization algorithm [1], I
couldn't do a simple optimization because of the limitations of the
iteration API [2]. Even if we pushed all the development effort to make
the batch API more suitable for ML there would be things we couldn't do.
E.g. there are approaches for updating a model iteratively without locks
[3,4] (i.e. somewhat asynchronously), and I don't see a clear way to
implement such algorithms with the batch API.

3) Streaming community (users and devs) benefit
The Flink streaming community in general would also benefit from this
direction. There are many features needed in the streaming API for ML to
work, but this is also true for the batch API. One really important is
the loops API (a.k.a. iterative DataStreams) [5]. There has been a lot
of effort (mostly from Paris) for making it mature enough [6]. Kate
mentioned using GPUs, and I'm sure they have uses in streaming generally
[7]. Thus, by improving the streaming API to allow ML algorithms, the
streaming API benefit too (which is important as they have a lot more
production users than the batch API).

4) Performance can be at least as good
I believe the same performance could be achieved with the streaming API
as with the batch API. Streaming API is much closer to the runtime than
the batch API. For corner-cases, with runtime-layer optimizations of
batch API, we could find a way to do the same (or similar) optimization
for the streaming API (see my previous point). Such case could be using
managed memory (and spilling to disk). There are also benefits by
default, e.g. we would have a finer grained fault tolerance with the
streaming API.

5) We could keep batch ML API
For the shorter term, we should not throw away all the algorithms
implemented with the batch API. By pushing forward the development with
side inputs we could make them usable with streaming API. Then, if the
library gains some popularity, we could replace the algorithms in the
batch API with streaming ones, to avoid the performance costs of e.g.
not being able to persist.

6) General tools for implementing ML algorithms
Besides implementing algorithms one by one, we could give more general
tools for making it easier to implement algorithms. E.g. parameter
server [8,9]. Theo also mentioned in another thread that TensorFlow has
a similar model to Flink streaming, we could look into that too. I think
often when deploying a production ML system, much more configuration and
tweaking should be done than e.g. Spark MLlib allows. Why not allow that?

7) Showcasing
Showcasing this could be easier. We could say that we're doing batch ML
with a streaming API. That's interesting in its own. IMHO this
integration is also a more approachable way towards end-to-end ML.


Thanks for reading so far :)

[1] https://github.com/apache/flink/pull/2819
[2] https://issues.apache.org/jira/browse/FLINK-2396
[3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
[4]
https://www.usenix.org/system/files/conference/hotos13/hotos13-final77.pdf
[5]
https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+Scoped+Loops+and+Job+Termination 

[6] https://github.com/apache/flink/pull/1668
[7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
[8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
[9]
http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/Using-QueryableState-inside-Flink-jobs-and-Parameter-Server-implementation-td15880.html

Cheers,
Gabor

[Theodore Vasiloudis]

Hello all,


@Gabor, we have discussed the idea of using the streaming API to write all
of our ML algorithms with a couple of people offline,
and I think it might be possible and is generally worth a shot. The
approach we would take would be close to Vowpal Wabbit, not exactly
"online", but rather "fast-batch".

There will be problems popping up again, even for very simple algos like on
line linear regression with SGD [1], but hopefully fixing those will be
more aligned with the priorities of the community.

@Katherin, my understanding is that given the limited resources, there is
no development effort focused on batch processing right now.

So to summarize, it seems like there are people willing to work on ML on
Flink, but nobody is sure how to do it.
There are many directions we could take (batch, online, batch on
streaming), each with its own merits and downsides.

If you want we can start a design doc and move the conversation there, come
up with a roadmap and start implementing.

Regards,
Theodore

[1]
http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/Understanding-connected-streams-use-without-timestamps-td10241.html

On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <[hidden email]>
wrote:

> It's great to see so much activity in this discussion :)
> I'll try to add my thoughts.
>
> I think building a developer community (Till's 2. point) can be slightly
> separated from what features we should aim for (1. point) and showcasing
> (3. point). Thanks Till for bringing up the ideas for restructuring, I'm
> sure we'll find a way to make the development process more dynamic. I'll
> try to address the rest here.
>
> It's hard to choose directions between streaming and batch ML. As Theo has
> indicated, not much online ML is used in production, but Flink concentrates
> on streaming, so online ML would be a better fit for Flink. However, as
> most of you argued, there's definite need for batch ML. But batch ML seems
> hard to achieve because there are blocking issues with persisting,
> iteration paths etc. So it's no good either way.
>
> I propose a seemingly crazy solution: what if we developed batch
> algorithms also with the streaming API? The batch API would clearly seem
> more suitable for ML algorithms, but there a lot of benefits of this
> approach too, so it's clearly worth considering. Flink also has the high
> level vision of "streaming for everything" that would clearly fit this
> case. What do you all think about this? Do you think this solution would be
> feasible? I would be happy to make a more elaborate proposal, but I push my
> main ideas here:
>
> 1) Simplifying by using one system
> It could simplify the work of both the users and the developers. One could
> execute training once, or could execute it periodically e.g. by using
> windows. Low-latency serving and training could be done in the same system.
> We could implement incremental algorithms, without any side inputs for
> combining online learning (or predictions) with batch learning. Of course,
> all the logic describing these must be somehow implemented (e.g.
> synchronizing predictions with training), but it should be easier to do so
> in one system, than by combining e.g. the batch and streaming API.
>
> 2) Batch ML with the streaming API is not harder
> Despite these benefits, it could seem harder to implement batch ML with
> the streaming API, but in my opinion it's not. There are more flexible,
> lower-level optimization potentials with the streaming API. Most
> distributed ML algorithms use a lower-level model than the batch API
> anyway, so sometimes it feels like forcing the algorithm logic into the
> training API and tweaking it. Although we could not use the batch
> primitives like join, we would have the E.g. in my experience with
> implementing a distributed matrix factorization algorithm [1], I couldn't
> do a simple optimization because of the limitations of the iteration API
> [2]. Even if we pushed all the development effort to make the batch API
> more suitable for ML there would be things we couldn't do. E.g. there are
> approaches for updating a model iteratively without locks [3,4] (i.e.
> somewhat asynchronously), and I don't see a clear way to implement such
> algorithms with the batch API.
>
> 3) Streaming community (users and devs) benefit
> The Flink streaming community in general would also benefit from this
> direction. There are many features needed in the streaming API for ML to
> work, but this is also true for the batch API. One really important is the
> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot of
> effort (mostly from Paris) for making it mature enough [6]. Kate mentioned
> using GPUs, and I'm sure they have uses in streaming generally [7]. Thus,
> by improving the streaming API to allow ML algorithms, the streaming API
> benefit too (which is important as they have a lot more production users
> than the batch API).
>
> 4) Performance can be at least as good
> I believe the same performance could be achieved with the streaming API as
> with the batch API. Streaming API is much closer to the runtime than the
> batch API. For corner-cases, with runtime-layer optimizations of batch API,
> we could find a way to do the same (or similar) optimization for the
> streaming API (see my previous point). Such case could be using managed
> memory (and spilling to disk). There are also benefits by default, e.g. we
> would have a finer grained fault tolerance with the streaming API.
>
> 5) We could keep batch ML API
> For the shorter term, we should not throw away all the algorithms
> implemented with the batch API. By pushing forward the development with
> side inputs we could make them usable with streaming API. Then, if the
> library gains some popularity, we could replace the algorithms in the batch
> API with streaming ones, to avoid the performance costs of e.g. not being
> able to persist.
>
> 6) General tools for implementing ML algorithms
> Besides implementing algorithms one by one, we could give more general
> tools for making it easier to implement algorithms. E.g. parameter server
> [8,9]. Theo also mentioned in another thread that TensorFlow has a similar
> model to Flink streaming, we could look into that too. I think often when
> deploying a production ML system, much more configuration and tweaking
> should be done than e.g. Spark MLlib allows. Why not allow that?
>
> 7) Showcasing
> Showcasing this could be easier. We could say that we're doing batch ML
> with a streaming API. That's interesting in its own. IMHO this integration
> is also a more approachable way towards end-to-end ML.
>
>
> Thanks for reading so far :)
>
> [1] https://github.com/apache/flink/pull/2819
> [2] https://issues.apache.org/jira/browse/FLINK-2396
> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
> 13-final77.pdf
> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
> Scoped+Loops+and+Job+Termination
> [6] https://github.com/apache/flink/pull/1668
> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> com/Using-QueryableState-inside-Flink-jobs-and-
> Parameter-Server-implementation-td15880.html
>
> Cheers,
> Gabor
>

[Gábor Hermann]

@Theodore: Great to hear you think the "batch on streaming" approach is
possible! Of course, we need to pay attention all the pitfalls there, if
we go that way.

+1 for a design doc!

I would add that it's possible to make efforts in all the three
directions (i.e. batch, online, batch on streaming) at the same time.
Although, it might be worth to concentrate on one. E.g. it would not be
so useful to have the same batch algorithms with both the batch API and
streaming API. We can decide later.

The design doc could be partitioned to these 3 directions, and we can
collect there the pros/cons too. What do you think?

Cheers,
Gabor


On 2017-02-23 12:13, Theodore Vasiloudis wrote:

> Hello all,
>
>
> @Gabor, we have discussed the idea of using the streaming API to write all
> of our ML algorithms with a couple of people offline,
> and I think it might be possible and is generally worth a shot. The
> approach we would take would be close to Vowpal Wabbit, not exactly
> "online", but rather "fast-batch".
>
> There will be problems popping up again, even for very simple algos like on
> line linear regression with SGD [1], but hopefully fixing those will be
> more aligned with the priorities of the community.
>
> @Katherin, my understanding is that given the limited resources, there is
> no development effort focused on batch processing right now.
>
> So to summarize, it seems like there are people willing to work on ML on
> Flink, but nobody is sure how to do it.
> There are many directions we could take (batch, online, batch on
> streaming), each with its own merits and downsides.
>
> If you want we can start a design doc and move the conversation there, come
> up with a roadmap and start implementing.
>
> Regards,
> Theodore
>
> [1]
> http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/Understanding-connected-streams-use-without-timestamps-td10241.html
>
> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <[hidden email]>
> wrote:
>
>> It's great to see so much activity in this discussion :)
>> I'll try to add my thoughts.
>>
>> I think building a developer community (Till's 2. point) can be slightly
>> separated from what features we should aim for (1. point) and showcasing
>> (3. point). Thanks Till for bringing up the ideas for restructuring, I'm
>> sure we'll find a way to make the development process more dynamic. I'll
>> try to address the rest here.
>>
>> It's hard to choose directions between streaming and batch ML. As Theo has
>> indicated, not much online ML is used in production, but Flink concentrates
>> on streaming, so online ML would be a better fit for Flink. However, as
>> most of you argued, there's definite need for batch ML. But batch ML seems
>> hard to achieve because there are blocking issues with persisting,
>> iteration paths etc. So it's no good either way.
>>
>> I propose a seemingly crazy solution: what if we developed batch
>> algorithms also with the streaming API? The batch API would clearly seem
>> more suitable for ML algorithms, but there a lot of benefits of this
>> approach too, so it's clearly worth considering. Flink also has the high
>> level vision of "streaming for everything" that would clearly fit this
>> case. What do you all think about this? Do you think this solution would be
>> feasible? I would be happy to make a more elaborate proposal, but I push my
>> main ideas here:
>>
>> 1) Simplifying by using one system
>> It could simplify the work of both the users and the developers. One could
>> execute training once, or could execute it periodically e.g. by using
>> windows. Low-latency serving and training could be done in the same system.
>> We could implement incremental algorithms, without any side inputs for
>> combining online learning (or predictions) with batch learning. Of course,
>> all the logic describing these must be somehow implemented (e.g.
>> synchronizing predictions with training), but it should be easier to do so
>> in one system, than by combining e.g. the batch and streaming API.
>>
>> 2) Batch ML with the streaming API is not harder
>> Despite these benefits, it could seem harder to implement batch ML with
>> the streaming API, but in my opinion it's not. There are more flexible,
>> lower-level optimization potentials with the streaming API. Most
>> distributed ML algorithms use a lower-level model than the batch API
>> anyway, so sometimes it feels like forcing the algorithm logic into the
>> training API and tweaking it. Although we could not use the batch
>> primitives like join, we would have the E.g. in my experience with
>> implementing a distributed matrix factorization algorithm [1], I couldn't
>> do a simple optimization because of the limitations of the iteration API
>> [2]. Even if we pushed all the development effort to make the batch API
>> more suitable for ML there would be things we couldn't do. E.g. there are
>> approaches for updating a model iteratively without locks [3,4] (i.e.
>> somewhat asynchronously), and I don't see a clear way to implement such
>> algorithms with the batch API.
>>
>> 3) Streaming community (users and devs) benefit
>> The Flink streaming community in general would also benefit from this
>> direction. There are many features needed in the streaming API for ML to
>> work, but this is also true for the batch API. One really important is the
>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot of
>> effort (mostly from Paris) for making it mature enough [6]. Kate mentioned
>> using GPUs, and I'm sure they have uses in streaming generally [7]. Thus,
>> by improving the streaming API to allow ML algorithms, the streaming API
>> benefit too (which is important as they have a lot more production users
>> than the batch API).
>>
>> 4) Performance can be at least as good
>> I believe the same performance could be achieved with the streaming API as
>> with the batch API. Streaming API is much closer to the runtime than the
>> batch API. For corner-cases, with runtime-layer optimizations of batch API,
>> we could find a way to do the same (or similar) optimization for the
>> streaming API (see my previous point). Such case could be using managed
>> memory (and spilling to disk). There are also benefits by default, e.g. we
>> would have a finer grained fault tolerance with the streaming API.
>>
>> 5) We could keep batch ML API
>> For the shorter term, we should not throw away all the algorithms
>> implemented with the batch API. By pushing forward the development with
>> side inputs we could make them usable with streaming API. Then, if the
>> library gains some popularity, we could replace the algorithms in the batch
>> API with streaming ones, to avoid the performance costs of e.g. not being
>> able to persist.
>>
>> 6) General tools for implementing ML algorithms
>> Besides implementing algorithms one by one, we could give more general
>> tools for making it easier to implement algorithms. E.g. parameter server
>> [8,9]. Theo also mentioned in another thread that TensorFlow has a similar
>> model to Flink streaming, we could look into that too. I think often when
>> deploying a production ML system, much more configuration and tweaking
>> should be done than e.g. Spark MLlib allows. Why not allow that?
>>
>> 7) Showcasing
>> Showcasing this could be easier. We could say that we're doing batch ML
>> with a streaming API. That's interesting in its own. IMHO this integration
>> is also a more approachable way towards end-to-end ML.
>>
>>
>> Thanks for reading so far :)
>>
>> [1] https://github.com/apache/flink/pull/2819
>> [2] https://issues.apache.org/jira/browse/FLINK-2396
>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
>> 13-final77.pdf
>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
>> Scoped+Loops+and+Job+Termination
>> [6] https://github.com/apache/flink/pull/1668
>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
>> com/Using-QueryableState-inside-Flink-jobs-and-
>> Parameter-Server-implementation-td15880.html
>>
>> Cheers,
>> Gabor
>>

[Katherin Eri]

I'm not sure that this is feasible, doing all at the same time could mean
doing nothing((((
I'm just afraid, that words: we will work on streaming not on batching, we
have no commiter's time for this, mean that yes, we started work on
FLINK-1730, but nobody will commit this work in the end, as it already was
with this ticket.

23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <[hidden email]>
написал:

> @Theodore: Great to hear you think the "batch on streaming" approach is
> possible! Of course, we need to pay attention all the pitfalls there, if we
> go that way.
>
> +1 for a design doc!
>
> I would add that it's possible to make efforts in all the three directions
> (i.e. batch, online, batch on streaming) at the same time. Although, it
> might be worth to concentrate on one. E.g. it would not be so useful to
> have the same batch algorithms with both the batch API and streaming API.
> We can decide later.
>
> The design doc could be partitioned to these 3 directions, and we can
> collect there the pros/cons too. What do you think?
>
> Cheers,
> Gabor
>
>
> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
>
>> Hello all,
>>
>>
>> @Gabor, we have discussed the idea of using the streaming API to write all
>> of our ML algorithms with a couple of people offline,
>> and I think it might be possible and is generally worth a shot. The
>> approach we would take would be close to Vowpal Wabbit, not exactly
>> "online", but rather "fast-batch".
>>
>> There will be problems popping up again, even for very simple algos like
>> on
>> line linear regression with SGD [1], but hopefully fixing those will be
>> more aligned with the priorities of the community.
>>
>> @Katherin, my understanding is that given the limited resources, there is
>> no development effort focused on batch processing right now.
>>
>> So to summarize, it seems like there are people willing to work on ML on
>> Flink, but nobody is sure how to do it.
>> There are many directions we could take (batch, online, batch on
>> streaming), each with its own merits and downsides.
>>
>> If you want we can start a design doc and move the conversation there,
>> come
>> up with a roadmap and start implementing.
>>
>> Regards,
>> Theodore
>>
>> [1]
>> http://apache-flink-user-mailing-list-archive.2336050.n4.
>> nabble.com/Understanding-connected-streams-use-without-times
>> tamps-td10241.html
>>
>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <[hidden email]>
>> wrote:
>>
>> It's great to see so much activity in this discussion :)
>>> I'll try to add my thoughts.
>>>
>>> I think building a developer community (Till's 2. point) can be slightly
>>> separated from what features we should aim for (1. point) and showcasing
>>> (3. point). Thanks Till for bringing up the ideas for restructuring, I'm
>>> sure we'll find a way to make the development process more dynamic. I'll
>>> try to address the rest here.
>>>
>>> It's hard to choose directions between streaming and batch ML. As Theo
>>> has
>>> indicated, not much online ML is used in production, but Flink
>>> concentrates
>>> on streaming, so online ML would be a better fit for Flink. However, as
>>> most of you argued, there's definite need for batch ML. But batch ML
>>> seems
>>> hard to achieve because there are blocking issues with persisting,
>>> iteration paths etc. So it's no good either way.
>>>
>>> I propose a seemingly crazy solution: what if we developed batch
>>> algorithms also with the streaming API? The batch API would clearly seem
>>> more suitable for ML algorithms, but there a lot of benefits of this
>>> approach too, so it's clearly worth considering. Flink also has the high
>>> level vision of "streaming for everything" that would clearly fit this
>>> case. What do you all think about this? Do you think this solution would
>>> be
>>> feasible? I would be happy to make a more elaborate proposal, but I push
>>> my
>>> main ideas here:
>>>
>>> 1) Simplifying by using one system
>>> It could simplify the work of both the users and the developers. One
>>> could
>>> execute training once, or could execute it periodically e.g. by using
>>> windows. Low-latency serving and training could be done in the same
>>> system.
>>> We could implement incremental algorithms, without any side inputs for
>>> combining online learning (or predictions) with batch learning. Of
>>> course,
>>> all the logic describing these must be somehow implemented (e.g.
>>> synchronizing predictions with training), but it should be easier to do
>>> so
>>> in one system, than by combining e.g. the batch and streaming API.
>>>
>>> 2) Batch ML with the streaming API is not harder
>>> Despite these benefits, it could seem harder to implement batch ML with
>>> the streaming API, but in my opinion it's not. There are more flexible,
>>> lower-level optimization potentials with the streaming API. Most
>>> distributed ML algorithms use a lower-level model than the batch API
>>> anyway, so sometimes it feels like forcing the algorithm logic into the
>>> training API and tweaking it. Although we could not use the batch
>>> primitives like join, we would have the E.g. in my experience with
>>> implementing a distributed matrix factorization algorithm [1], I couldn't
>>> do a simple optimization because of the limitations of the iteration API
>>> [2]. Even if we pushed all the development effort to make the batch API
>>> more suitable for ML there would be things we couldn't do. E.g. there are
>>> approaches for updating a model iteratively without locks [3,4] (i.e.
>>> somewhat asynchronously), and I don't see a clear way to implement such
>>> algorithms with the batch API.
>>>
>>> 3) Streaming community (users and devs) benefit
>>> The Flink streaming community in general would also benefit from this
>>> direction. There are many features needed in the streaming API for ML to
>>> work, but this is also true for the batch API. One really important is
>>> the
>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot of
>>> effort (mostly from Paris) for making it mature enough [6]. Kate
>>> mentioned
>>> using GPUs, and I'm sure they have uses in streaming generally [7]. Thus,
>>> by improving the streaming API to allow ML algorithms, the streaming API
>>> benefit too (which is important as they have a lot more production users
>>> than the batch API).
>>>
>>> 4) Performance can be at least as good
>>> I believe the same performance could be achieved with the streaming API
>>> as
>>> with the batch API. Streaming API is much closer to the runtime than the
>>> batch API. For corner-cases, with runtime-layer optimizations of batch
>>> API,
>>> we could find a way to do the same (or similar) optimization for the
>>> streaming API (see my previous point). Such case could be using managed
>>> memory (and spilling to disk). There are also benefits by default, e.g.
>>> we
>>> would have a finer grained fault tolerance with the streaming API.
>>>
>>> 5) We could keep batch ML API
>>> For the shorter term, we should not throw away all the algorithms
>>> implemented with the batch API. By pushing forward the development with
>>> side inputs we could make them usable with streaming API. Then, if the
>>> library gains some popularity, we could replace the algorithms in the
>>> batch
>>> API with streaming ones, to avoid the performance costs of e.g. not being
>>> able to persist.
>>>
>>> 6) General tools for implementing ML algorithms
>>> Besides implementing algorithms one by one, we could give more general
>>> tools for making it easier to implement algorithms. E.g. parameter server
>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
>>> similar
>>> model to Flink streaming, we could look into that too. I think often when
>>> deploying a production ML system, much more configuration and tweaking
>>> should be done than e.g. Spark MLlib allows. Why not allow that?
>>>
>>> 7) Showcasing
>>> Showcasing this could be easier. We could say that we're doing batch ML
>>> with a streaming API. That's interesting in its own. IMHO this
>>> integration
>>> is also a more approachable way towards end-to-end ML.
>>>
>>>
>>> Thanks for reading so far :)
>>>
>>> [1] https://github.com/apache/flink/pull/2819
>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
>>> 13-final77.pdf
>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
>>> Scoped+Loops+and+Job+Termination
>>> [6] https://github.com/apache/flink/pull/1668
>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
>>> com/Using-QueryableState-inside-Flink-jobs-and-
>>> Parameter-Server-implementation-td15880.html
>>>
>>> Cheers,
>>> Gabor
>>>
>>>
>

[Gábor Hermann]

I agree, that it's better to go in one direction first, but I think
online and offline with streaming API can go somewhat parallel later. We
could set a short-term goal, concentrate initially on one direction, and
showcase that direction (e.g. in a blogpost). But first, we should list
the pros/cons in a design doc as a minimum. Then make a decision what
direction to go. Would that be feasible?

On 2017-02-23 12:34, Katherin Eri wrote:

> I'm not sure that this is feasible, doing all at the same time could mean
> doing nothing((((
> I'm just afraid, that words: we will work on streaming not on batching, we
> have no commiter's time for this, mean that yes, we started work on
> FLINK-1730, but nobody will commit this work in the end, as it already was
> with this ticket.
>
> 23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <[hidden email]>
> написал:
>
>> @Theodore: Great to hear you think the "batch on streaming" approach is
>> possible! Of course, we need to pay attention all the pitfalls there, if we
>> go that way.
>>
>> +1 for a design doc!
>>
>> I would add that it's possible to make efforts in all the three directions
>> (i.e. batch, online, batch on streaming) at the same time. Although, it
>> might be worth to concentrate on one. E.g. it would not be so useful to
>> have the same batch algorithms with both the batch API and streaming API.
>> We can decide later.
>>
>> The design doc could be partitioned to these 3 directions, and we can
>> collect there the pros/cons too. What do you think?
>>
>> Cheers,
>> Gabor
>>
>>
>> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
>>
>>> Hello all,
>>>
>>>
>>> @Gabor, we have discussed the idea of using the streaming API to write all
>>> of our ML algorithms with a couple of people offline,
>>> and I think it might be possible and is generally worth a shot. The
>>> approach we would take would be close to Vowpal Wabbit, not exactly
>>> "online", but rather "fast-batch".
>>>
>>> There will be problems popping up again, even for very simple algos like
>>> on
>>> line linear regression with SGD [1], but hopefully fixing those will be
>>> more aligned with the priorities of the community.
>>>
>>> @Katherin, my understanding is that given the limited resources, there is
>>> no development effort focused on batch processing right now.
>>>
>>> So to summarize, it seems like there are people willing to work on ML on
>>> Flink, but nobody is sure how to do it.
>>> There are many directions we could take (batch, online, batch on
>>> streaming), each with its own merits and downsides.
>>>
>>> If you want we can start a design doc and move the conversation there,
>>> come
>>> up with a roadmap and start implementing.
>>>
>>> Regards,
>>> Theodore
>>>
>>> [1]
>>> http://apache-flink-user-mailing-list-archive.2336050.n4.
>>> nabble.com/Understanding-connected-streams-use-without-times
>>> tamps-td10241.html
>>>
>>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <[hidden email]>
>>> wrote:
>>>
>>> It's great to see so much activity in this discussion :)
>>>> I'll try to add my thoughts.
>>>>
>>>> I think building a developer community (Till's 2. point) can be slightly
>>>> separated from what features we should aim for (1. point) and showcasing
>>>> (3. point). Thanks Till for bringing up the ideas for restructuring, I'm
>>>> sure we'll find a way to make the development process more dynamic. I'll
>>>> try to address the rest here.
>>>>
>>>> It's hard to choose directions between streaming and batch ML. As Theo
>>>> has
>>>> indicated, not much online ML is used in production, but Flink
>>>> concentrates
>>>> on streaming, so online ML would be a better fit for Flink. However, as
>>>> most of you argued, there's definite need for batch ML. But batch ML
>>>> seems
>>>> hard to achieve because there are blocking issues with persisting,
>>>> iteration paths etc. So it's no good either way.
>>>>
>>>> I propose a seemingly crazy solution: what if we developed batch
>>>> algorithms also with the streaming API? The batch API would clearly seem
>>>> more suitable for ML algorithms, but there a lot of benefits of this
>>>> approach too, so it's clearly worth considering. Flink also has the high
>>>> level vision of "streaming for everything" that would clearly fit this
>>>> case. What do you all think about this? Do you think this solution would
>>>> be
>>>> feasible? I would be happy to make a more elaborate proposal, but I push
>>>> my
>>>> main ideas here:
>>>>
>>>> 1) Simplifying by using one system
>>>> It could simplify the work of both the users and the developers. One
>>>> could
>>>> execute training once, or could execute it periodically e.g. by using
>>>> windows. Low-latency serving and training could be done in the same
>>>> system.
>>>> We could implement incremental algorithms, without any side inputs for
>>>> combining online learning (or predictions) with batch learning. Of
>>>> course,
>>>> all the logic describing these must be somehow implemented (e.g.
>>>> synchronizing predictions with training), but it should be easier to do
>>>> so
>>>> in one system, than by combining e.g. the batch and streaming API.
>>>>
>>>> 2) Batch ML with the streaming API is not harder
>>>> Despite these benefits, it could seem harder to implement batch ML with
>>>> the streaming API, but in my opinion it's not. There are more flexible,
>>>> lower-level optimization potentials with the streaming API. Most
>>>> distributed ML algorithms use a lower-level model than the batch API
>>>> anyway, so sometimes it feels like forcing the algorithm logic into the
>>>> training API and tweaking it. Although we could not use the batch
>>>> primitives like join, we would have the E.g. in my experience with
>>>> implementing a distributed matrix factorization algorithm [1], I couldn't
>>>> do a simple optimization because of the limitations of the iteration API
>>>> [2]. Even if we pushed all the development effort to make the batch API
>>>> more suitable for ML there would be things we couldn't do. E.g. there are
>>>> approaches for updating a model iteratively without locks [3,4] (i.e.
>>>> somewhat asynchronously), and I don't see a clear way to implement such
>>>> algorithms with the batch API.
>>>>
>>>> 3) Streaming community (users and devs) benefit
>>>> The Flink streaming community in general would also benefit from this
>>>> direction. There are many features needed in the streaming API for ML to
>>>> work, but this is also true for the batch API. One really important is
>>>> the
>>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot of
>>>> effort (mostly from Paris) for making it mature enough [6]. Kate
>>>> mentioned
>>>> using GPUs, and I'm sure they have uses in streaming generally [7]. Thus,
>>>> by improving the streaming API to allow ML algorithms, the streaming API
>>>> benefit too (which is important as they have a lot more production users
>>>> than the batch API).
>>>>
>>>> 4) Performance can be at least as good
>>>> I believe the same performance could be achieved with the streaming API
>>>> as
>>>> with the batch API. Streaming API is much closer to the runtime than the
>>>> batch API. For corner-cases, with runtime-layer optimizations of batch
>>>> API,
>>>> we could find a way to do the same (or similar) optimization for the
>>>> streaming API (see my previous point). Such case could be using managed
>>>> memory (and spilling to disk). There are also benefits by default, e.g.
>>>> we
>>>> would have a finer grained fault tolerance with the streaming API.
>>>>
>>>> 5) We could keep batch ML API
>>>> For the shorter term, we should not throw away all the algorithms
>>>> implemented with the batch API. By pushing forward the development with
>>>> side inputs we could make them usable with streaming API. Then, if the
>>>> library gains some popularity, we could replace the algorithms in the
>>>> batch
>>>> API with streaming ones, to avoid the performance costs of e.g. not being
>>>> able to persist.
>>>>
>>>> 6) General tools for implementing ML algorithms
>>>> Besides implementing algorithms one by one, we could give more general
>>>> tools for making it easier to implement algorithms. E.g. parameter server
>>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
>>>> similar
>>>> model to Flink streaming, we could look into that too. I think often when
>>>> deploying a production ML system, much more configuration and tweaking
>>>> should be done than e.g. Spark MLlib allows. Why not allow that?
>>>>
>>>> 7) Showcasing
>>>> Showcasing this could be easier. We could say that we're doing batch ML
>>>> with a streaming API. That's interesting in its own. IMHO this
>>>> integration
>>>> is also a more approachable way towards end-to-end ML.
>>>>
>>>>
>>>> Thanks for reading so far :)
>>>>
>>>> [1] https://github.com/apache/flink/pull/2819
>>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
>>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
>>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
>>>> 13-final77.pdf
>>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
>>>> Scoped+Loops+and+Job+Termination
>>>> [6] https://github.com/apache/flink/pull/1668
>>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
>>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
>>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
>>>> com/Using-QueryableState-inside-Flink-jobs-and-
>>>> Parameter-Server-implementation-td15880.html
>>>>
>>>> Cheers,
>>>> Gabor
>>>>
>>>>

[Katherin Eri]

Yes, ok.
let's start some design document, and write down there already mentioned
ideas about: parameter server, about clipper and others. Would be nice if
we will also map this approaches to cases.
Will work on it collaboratively on each topic, may be finally we will form
some picture, that could be agreed with committers.
@Gabor, could you please start such shared doc, as you have already several
ideas proposed?

чт, 23 февр. 2017, 15:06 Gábor Hermann <[hidden email]>:

> I agree, that it's better to go in one direction first, but I think
> online and offline with streaming API can go somewhat parallel later. We
> could set a short-term goal, concentrate initially on one direction, and
> showcase that direction (e.g. in a blogpost). But first, we should list
> the pros/cons in a design doc as a minimum. Then make a decision what
> direction to go. Would that be feasible?
>
> On 2017-02-23 12:34, Katherin Eri wrote:
>
> > I'm not sure that this is feasible, doing all at the same time could mean
> > doing nothing((((
> > I'm just afraid, that words: we will work on streaming not on batching,
> we
> > have no commiter's time for this, mean that yes, we started work on
> > FLINK-1730, but nobody will commit this work in the end, as it already
> was
> > with this ticket.
> >
> > 23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <
> [hidden email]>
> > написал:
> >
> >> @Theodore: Great to hear you think the "batch on streaming" approach is
> >> possible! Of course, we need to pay attention all the pitfalls there,
> if we
> >> go that way.
> >>
> >> +1 for a design doc!
> >>
> >> I would add that it's possible to make efforts in all the three
> directions
> >> (i.e. batch, online, batch on streaming) at the same time. Although, it
> >> might be worth to concentrate on one. E.g. it would not be so useful to
> >> have the same batch algorithms with both the batch API and streaming
> API.
> >> We can decide later.
> >>
> >> The design doc could be partitioned to these 3 directions, and we can
> >> collect there the pros/cons too. What do you think?
> >>
> >> Cheers,
> >> Gabor
> >>
> >>
> >> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
> >>
> >>> Hello all,
> >>>
> >>>
> >>> @Gabor, we have discussed the idea of using the streaming API to write
> all
> >>> of our ML algorithms with a couple of people offline,
> >>> and I think it might be possible and is generally worth a shot. The
> >>> approach we would take would be close to Vowpal Wabbit, not exactly
> >>> "online", but rather "fast-batch".
> >>>
> >>> There will be problems popping up again, even for very simple algos
> like
> >>> on
> >>> line linear regression with SGD [1], but hopefully fixing those will be
> >>> more aligned with the priorities of the community.
> >>>
> >>> @Katherin, my understanding is that given the limited resources, there
> is
> >>> no development effort focused on batch processing right now.
> >>>
> >>> So to summarize, it seems like there are people willing to work on ML
> on
> >>> Flink, but nobody is sure how to do it.
> >>> There are many directions we could take (batch, online, batch on
> >>> streaming), each with its own merits and downsides.
> >>>
> >>> If you want we can start a design doc and move the conversation there,
> >>> come
> >>> up with a roadmap and start implementing.
> >>>
> >>> Regards,
> >>> Theodore
> >>>
> >>> [1]
> >>> http://apache-flink-user-mailing-list-archive.2336050.n4.
> >>> nabble.com/Understanding-connected-streams-use-without-times
> >>> tamps-td10241.html
> >>>
> >>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <[hidden email]
> >
> >>> wrote:
> >>>
> >>> It's great to see so much activity in this discussion :)
> >>>> I'll try to add my thoughts.
> >>>>
> >>>> I think building a developer community (Till's 2. point) can be
> slightly
> >>>> separated from what features we should aim for (1. point) and
> showcasing
> >>>> (3. point). Thanks Till for bringing up the ideas for restructuring,
> I'm
> >>>> sure we'll find a way to make the development process more dynamic.
> I'll
> >>>> try to address the rest here.
> >>>>
> >>>> It's hard to choose directions between streaming and batch ML. As Theo
> >>>> has
> >>>> indicated, not much online ML is used in production, but Flink
> >>>> concentrates
> >>>> on streaming, so online ML would be a better fit for Flink. However,
> as
> >>>> most of you argued, there's definite need for batch ML. But batch ML
> >>>> seems
> >>>> hard to achieve because there are blocking issues with persisting,
> >>>> iteration paths etc. So it's no good either way.
> >>>>
> >>>> I propose a seemingly crazy solution: what if we developed batch
> >>>> algorithms also with the streaming API? The batch API would clearly
> seem
> >>>> more suitable for ML algorithms, but there a lot of benefits of this
> >>>> approach too, so it's clearly worth considering. Flink also has the
> high
> >>>> level vision of "streaming for everything" that would clearly fit this
> >>>> case. What do you all think about this? Do you think this solution
> would
> >>>> be
> >>>> feasible? I would be happy to make a more elaborate proposal, but I
> push
> >>>> my
> >>>> main ideas here:
> >>>>
> >>>> 1) Simplifying by using one system
> >>>> It could simplify the work of both the users and the developers. One
> >>>> could
> >>>> execute training once, or could execute it periodically e.g. by using
> >>>> windows. Low-latency serving and training could be done in the same
> >>>> system.
> >>>> We could implement incremental algorithms, without any side inputs for
> >>>> combining online learning (or predictions) with batch learning. Of
> >>>> course,
> >>>> all the logic describing these must be somehow implemented (e.g.
> >>>> synchronizing predictions with training), but it should be easier to
> do
> >>>> so
> >>>> in one system, than by combining e.g. the batch and streaming API.
> >>>>
> >>>> 2) Batch ML with the streaming API is not harder
> >>>> Despite these benefits, it could seem harder to implement batch ML
> with
> >>>> the streaming API, but in my opinion it's not. There are more
> flexible,
> >>>> lower-level optimization potentials with the streaming API. Most
> >>>> distributed ML algorithms use a lower-level model than the batch API
> >>>> anyway, so sometimes it feels like forcing the algorithm logic into
> the
> >>>> training API and tweaking it. Although we could not use the batch
> >>>> primitives like join, we would have the E.g. in my experience with
> >>>> implementing a distributed matrix factorization algorithm [1], I
> couldn't
> >>>> do a simple optimization because of the limitations of the iteration
> API
> >>>> [2]. Even if we pushed all the development effort to make the batch
> API
> >>>> more suitable for ML there would be things we couldn't do. E.g. there
> are
> >>>> approaches for updating a model iteratively without locks [3,4] (i.e.
> >>>> somewhat asynchronously), and I don't see a clear way to implement
> such
> >>>> algorithms with the batch API.
> >>>>
> >>>> 3) Streaming community (users and devs) benefit
> >>>> The Flink streaming community in general would also benefit from this
> >>>> direction. There are many features needed in the streaming API for ML
> to
> >>>> work, but this is also true for the batch API. One really important is
> >>>> the
> >>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot of
> >>>> effort (mostly from Paris) for making it mature enough [6]. Kate
> >>>> mentioned
> >>>> using GPUs, and I'm sure they have uses in streaming generally [7].
> Thus,
> >>>> by improving the streaming API to allow ML algorithms, the streaming
> API
> >>>> benefit too (which is important as they have a lot more production
> users
> >>>> than the batch API).
> >>>>
> >>>> 4) Performance can be at least as good
> >>>> I believe the same performance could be achieved with the streaming
> API
> >>>> as
> >>>> with the batch API. Streaming API is much closer to the runtime than
> the
> >>>> batch API. For corner-cases, with runtime-layer optimizations of batch
> >>>> API,
> >>>> we could find a way to do the same (or similar) optimization for the
> >>>> streaming API (see my previous point). Such case could be using
> managed
> >>>> memory (and spilling to disk). There are also benefits by default,
> e.g.
> >>>> we
> >>>> would have a finer grained fault tolerance with the streaming API.
> >>>>
> >>>> 5) We could keep batch ML API
> >>>> For the shorter term, we should not throw away all the algorithms
> >>>> implemented with the batch API. By pushing forward the development
> with
> >>>> side inputs we could make them usable with streaming API. Then, if the
> >>>> library gains some popularity, we could replace the algorithms in the
> >>>> batch
> >>>> API with streaming ones, to avoid the performance costs of e.g. not
> being
> >>>> able to persist.
> >>>>
> >>>> 6) General tools for implementing ML algorithms
> >>>> Besides implementing algorithms one by one, we could give more general
> >>>> tools for making it easier to implement algorithms. E.g. parameter
> server
> >>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
> >>>> similar
> >>>> model to Flink streaming, we could look into that too. I think often
> when
> >>>> deploying a production ML system, much more configuration and tweaking
> >>>> should be done than e.g. Spark MLlib allows. Why not allow that?
> >>>>
> >>>> 7) Showcasing
> >>>> Showcasing this could be easier. We could say that we're doing batch
> ML
> >>>> with a streaming API. That's interesting in its own. IMHO this
> >>>> integration
> >>>> is also a more approachable way towards end-to-end ML.
> >>>>
> >>>>
> >>>> Thanks for reading so far :)
> >>>>
> >>>> [1] https://github.com/apache/flink/pull/2819
> >>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
> >>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
> >>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
> >>>> 13-final77.pdf
> >>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
> >>>> Scoped+Loops+and+Job+Termination
> >>>> [6] https://github.com/apache/flink/pull/1668
> >>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
> >>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
> >>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> >>>> com/Using-QueryableState-inside-Flink-jobs-and-
> >>>> Parameter-Server-implementation-td15880.html
> >>>>
> >>>> Cheers,
> >>>> Gabor
> >>>>
> >>>>
>
> --

*Yours faithfully, *

*Kate Eri.*

[Stavros Kontopoulos]

+100 for a design doc.

Could we also set a roadmap after some time-boxed investigation captured in
that document? We need action.

Looking forward to work on this (whatever that might be) ;) Also are there
any data supporting one direction or the other from a customer perspective?
It would help to make more informed decisions.

On Thu, Feb 23, 2017 at 2:23 PM, Katherin Eri <[hidden email]>
wrote:

> Yes, ok.
> let's start some design document, and write down there already mentioned
> ideas about: parameter server, about clipper and others. Would be nice if
> we will also map this approaches to cases.
> Will work on it collaboratively on each topic, may be finally we will form
> some picture, that could be agreed with committers.
> @Gabor, could you please start such shared doc, as you have already several
> ideas proposed?
>
> чт, 23 февр. 2017, 15:06 Gábor Hermann <[hidden email]>:
>
> > I agree, that it's better to go in one direction first, but I think
> > online and offline with streaming API can go somewhat parallel later. We
> > could set a short-term goal, concentrate initially on one direction, and
> > showcase that direction (e.g. in a blogpost). But first, we should list
> > the pros/cons in a design doc as a minimum. Then make a decision what
> > direction to go. Would that be feasible?
> >
> > On 2017-02-23 12:34, Katherin Eri wrote:
> >
> > > I'm not sure that this is feasible, doing all at the same time could
> mean
> > > doing nothing((((
> > > I'm just afraid, that words: we will work on streaming not on batching,
> > we
> > > have no commiter's time for this, mean that yes, we started work on
> > > FLINK-1730, but nobody will commit this work in the end, as it already
> > was
> > > with this ticket.
> > >
> > > 23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <
> > [hidden email]>
> > > написал:
> > >
> > >> @Theodore: Great to hear you think the "batch on streaming" approach
> is
> > >> possible! Of course, we need to pay attention all the pitfalls there,
> > if we
> > >> go that way.
> > >>
> > >> +1 for a design doc!
> > >>
> > >> I would add that it's possible to make efforts in all the three
> > directions
> > >> (i.e. batch, online, batch on streaming) at the same time. Although,
> it
> > >> might be worth to concentrate on one. E.g. it would not be so useful
> to
> > >> have the same batch algorithms with both the batch API and streaming
> > API.
> > >> We can decide later.
> > >>
> > >> The design doc could be partitioned to these 3 directions, and we can
> > >> collect there the pros/cons too. What do you think?
> > >>
> > >> Cheers,
> > >> Gabor
> > >>
> > >>
> > >> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
> > >>
> > >>> Hello all,
> > >>>
> > >>>
> > >>> @Gabor, we have discussed the idea of using the streaming API to
> write
> > all
> > >>> of our ML algorithms with a couple of people offline,
> > >>> and I think it might be possible and is generally worth a shot. The
> > >>> approach we would take would be close to Vowpal Wabbit, not exactly
> > >>> "online", but rather "fast-batch".
> > >>>
> > >>> There will be problems popping up again, even for very simple algos
> > like
> > >>> on
> > >>> line linear regression with SGD [1], but hopefully fixing those will
> be
> > >>> more aligned with the priorities of the community.
> > >>>
> > >>> @Katherin, my understanding is that given the limited resources,
> there
> > is
> > >>> no development effort focused on batch processing right now.
> > >>>
> > >>> So to summarize, it seems like there are people willing to work on ML
> > on
> > >>> Flink, but nobody is sure how to do it.
> > >>> There are many directions we could take (batch, online, batch on
> > >>> streaming), each with its own merits and downsides.
> > >>>
> > >>> If you want we can start a design doc and move the conversation
> there,
> > >>> come
> > >>> up with a roadmap and start implementing.
> > >>>
> > >>> Regards,
> > >>> Theodore
> > >>>
> > >>> [1]
> > >>> http://apache-flink-user-mailing-list-archive.2336050.n4.
> > >>> nabble.com/Understanding-connected-streams-use-without-times
> > >>> tamps-td10241.html
> > >>>
> > >>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <
> [hidden email]
> > >
> > >>> wrote:
> > >>>
> > >>> It's great to see so much activity in this discussion :)
> > >>>> I'll try to add my thoughts.
> > >>>>
> > >>>> I think building a developer community (Till's 2. point) can be
> > slightly
> > >>>> separated from what features we should aim for (1. point) and
> > showcasing
> > >>>> (3. point). Thanks Till for bringing up the ideas for restructuring,
> > I'm
> > >>>> sure we'll find a way to make the development process more dynamic.
> > I'll
> > >>>> try to address the rest here.
> > >>>>
> > >>>> It's hard to choose directions between streaming and batch ML. As
> Theo
> > >>>> has
> > >>>> indicated, not much online ML is used in production, but Flink
> > >>>> concentrates
> > >>>> on streaming, so online ML would be a better fit for Flink. However,
> > as
> > >>>> most of you argued, there's definite need for batch ML. But batch ML
> > >>>> seems
> > >>>> hard to achieve because there are blocking issues with persisting,
> > >>>> iteration paths etc. So it's no good either way.
> > >>>>
> > >>>> I propose a seemingly crazy solution: what if we developed batch
> > >>>> algorithms also with the streaming API? The batch API would clearly
> > seem
> > >>>> more suitable for ML algorithms, but there a lot of benefits of this
> > >>>> approach too, so it's clearly worth considering. Flink also has the
> > high
> > >>>> level vision of "streaming for everything" that would clearly fit
> this
> > >>>> case. What do you all think about this? Do you think this solution
> > would
> > >>>> be
> > >>>> feasible? I would be happy to make a more elaborate proposal, but I
> > push
> > >>>> my
> > >>>> main ideas here:
> > >>>>
> > >>>> 1) Simplifying by using one system
> > >>>> It could simplify the work of both the users and the developers. One
> > >>>> could
> > >>>> execute training once, or could execute it periodically e.g. by
> using
> > >>>> windows. Low-latency serving and training could be done in the same
> > >>>> system.
> > >>>> We could implement incremental algorithms, without any side inputs
> for
> > >>>> combining online learning (or predictions) with batch learning. Of
> > >>>> course,
> > >>>> all the logic describing these must be somehow implemented (e.g.
> > >>>> synchronizing predictions with training), but it should be easier to
> > do
> > >>>> so
> > >>>> in one system, than by combining e.g. the batch and streaming API.
> > >>>>
> > >>>> 2) Batch ML with the streaming API is not harder
> > >>>> Despite these benefits, it could seem harder to implement batch ML
> > with
> > >>>> the streaming API, but in my opinion it's not. There are more
> > flexible,
> > >>>> lower-level optimization potentials with the streaming API. Most
> > >>>> distributed ML algorithms use a lower-level model than the batch API
> > >>>> anyway, so sometimes it feels like forcing the algorithm logic into
> > the
> > >>>> training API and tweaking it. Although we could not use the batch
> > >>>> primitives like join, we would have the E.g. in my experience with
> > >>>> implementing a distributed matrix factorization algorithm [1], I
> > couldn't
> > >>>> do a simple optimization because of the limitations of the iteration
> > API
> > >>>> [2]. Even if we pushed all the development effort to make the batch
> > API
> > >>>> more suitable for ML there would be things we couldn't do. E.g.
> there
> > are
> > >>>> approaches for updating a model iteratively without locks [3,4]
> (i.e.
> > >>>> somewhat asynchronously), and I don't see a clear way to implement
> > such
> > >>>> algorithms with the batch API.
> > >>>>
> > >>>> 3) Streaming community (users and devs) benefit
> > >>>> The Flink streaming community in general would also benefit from
> this
> > >>>> direction. There are many features needed in the streaming API for
> ML
> > to
> > >>>> work, but this is also true for the batch API. One really important
> is
> > >>>> the
> > >>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot
> of
> > >>>> effort (mostly from Paris) for making it mature enough [6]. Kate
> > >>>> mentioned
> > >>>> using GPUs, and I'm sure they have uses in streaming generally [7].
> > Thus,
> > >>>> by improving the streaming API to allow ML algorithms, the streaming
> > API
> > >>>> benefit too (which is important as they have a lot more production
> > users
> > >>>> than the batch API).
> > >>>>
> > >>>> 4) Performance can be at least as good
> > >>>> I believe the same performance could be achieved with the streaming
> > API
> > >>>> as
> > >>>> with the batch API. Streaming API is much closer to the runtime than
> > the
> > >>>> batch API. For corner-cases, with runtime-layer optimizations of
> batch
> > >>>> API,
> > >>>> we could find a way to do the same (or similar) optimization for the
> > >>>> streaming API (see my previous point). Such case could be using
> > managed
> > >>>> memory (and spilling to disk). There are also benefits by default,
> > e.g.
> > >>>> we
> > >>>> would have a finer grained fault tolerance with the streaming API.
> > >>>>
> > >>>> 5) We could keep batch ML API
> > >>>> For the shorter term, we should not throw away all the algorithms
> > >>>> implemented with the batch API. By pushing forward the development
> > with
> > >>>> side inputs we could make them usable with streaming API. Then, if
> the
> > >>>> library gains some popularity, we could replace the algorithms in
> the
> > >>>> batch
> > >>>> API with streaming ones, to avoid the performance costs of e.g. not
> > being
> > >>>> able to persist.
> > >>>>
> > >>>> 6) General tools for implementing ML algorithms
> > >>>> Besides implementing algorithms one by one, we could give more
> general
> > >>>> tools for making it easier to implement algorithms. E.g. parameter
> > server
> > >>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
> > >>>> similar
> > >>>> model to Flink streaming, we could look into that too. I think often
> > when
> > >>>> deploying a production ML system, much more configuration and
> tweaking
> > >>>> should be done than e.g. Spark MLlib allows. Why not allow that?
> > >>>>
> > >>>> 7) Showcasing
> > >>>> Showcasing this could be easier. We could say that we're doing batch
> > ML
> > >>>> with a streaming API. That's interesting in its own. IMHO this
> > >>>> integration
> > >>>> is also a more approachable way towards end-to-end ML.
> > >>>>
> > >>>>
> > >>>> Thanks for reading so far :)
> > >>>>
> > >>>> [1] https://github.com/apache/flink/pull/2819
> > >>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
> > >>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
> > >>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
> > >>>> 13-final77.pdf
> > >>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
> > >>>> Scoped+Loops+and+Job+Termination
> > >>>> [6] https://github.com/apache/flink/pull/1668
> > >>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.pdf
> > >>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
> > >>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > >>>> com/Using-QueryableState-inside-Flink-jobs-and-
> > >>>> Parameter-Server-implementation-td15880.html
> > >>>>
> > >>>> Cheers,
> > >>>> Gabor
> > >>>>
> > >>>>
> >
> > --
>
> *Yours faithfully, *
>
> *Kate Eri.*
>

[Katherin Eri]

I have asked already some teams for useful cases, but all of them need time
to think.
During analysis something will finally arise.
May be we can ask partners of Flink  for cases? Data Artisans got results
of customers survey: [1], ML better support is wanted, so we could ask what
exactly is necessary.

[1] http://data-artisans.com/flink-user-survey-2016-part-2/

23 февр. 2017 г. 4:32 PM пользователь "Stavros Kontopoulos" <
[hidden email]> написал:

> +100 for a design doc.
>
> Could we also set a roadmap after some time-boxed investigation captured in
> that document? We need action.
>
> Looking forward to work on this (whatever that might be) ;) Also are there
> any data supporting one direction or the other from a customer perspective?
> It would help to make more informed decisions.
>
> On Thu, Feb 23, 2017 at 2:23 PM, Katherin Eri <[hidden email]>
> wrote:
>
> > Yes, ok.
> > let's start some design document, and write down there already mentioned
> > ideas about: parameter server, about clipper and others. Would be nice if
> > we will also map this approaches to cases.
> > Will work on it collaboratively on each topic, may be finally we will
> form
> > some picture, that could be agreed with committers.
> > @Gabor, could you please start such shared doc, as you have already
> several
> > ideas proposed?
> >
> > чт, 23 февр. 2017, 15:06 Gábor Hermann <[hidden email]>:
> >
> > > I agree, that it's better to go in one direction first, but I think
> > > online and offline with streaming API can go somewhat parallel later.
> We
> > > could set a short-term goal, concentrate initially on one direction,
> and
> > > showcase that direction (e.g. in a blogpost). But first, we should list
> > > the pros/cons in a design doc as a minimum. Then make a decision what
> > > direction to go. Would that be feasible?
> > >
> > > On 2017-02-23 12:34, Katherin Eri wrote:
> > >
> > > > I'm not sure that this is feasible, doing all at the same time could
> > mean
> > > > doing nothing((((
> > > > I'm just afraid, that words: we will work on streaming not on
> batching,
> > > we
> > > > have no commiter's time for this, mean that yes, we started work on
> > > > FLINK-1730, but nobody will commit this work in the end, as it
> already
> > > was
> > > > with this ticket.
> > > >
> > > > 23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <
> > > [hidden email]>
> > > > написал:
> > > >
> > > >> @Theodore: Great to hear you think the "batch on streaming" approach
> > is
> > > >> possible! Of course, we need to pay attention all the pitfalls
> there,
> > > if we
> > > >> go that way.
> > > >>
> > > >> +1 for a design doc!
> > > >>
> > > >> I would add that it's possible to make efforts in all the three
> > > directions
> > > >> (i.e. batch, online, batch on streaming) at the same time. Although,
> > it
> > > >> might be worth to concentrate on one. E.g. it would not be so useful
> > to
> > > >> have the same batch algorithms with both the batch API and streaming
> > > API.
> > > >> We can decide later.
> > > >>
> > > >> The design doc could be partitioned to these 3 directions, and we
> can
> > > >> collect there the pros/cons too. What do you think?
> > > >>
> > > >> Cheers,
> > > >> Gabor
> > > >>
> > > >>
> > > >> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
> > > >>
> > > >>> Hello all,
> > > >>>
> > > >>>
> > > >>> @Gabor, we have discussed the idea of using the streaming API to
> > write
> > > all
> > > >>> of our ML algorithms with a couple of people offline,
> > > >>> and I think it might be possible and is generally worth a shot. The
> > > >>> approach we would take would be close to Vowpal Wabbit, not exactly
> > > >>> "online", but rather "fast-batch".
> > > >>>
> > > >>> There will be problems popping up again, even for very simple algos
> > > like
> > > >>> on
> > > >>> line linear regression with SGD [1], but hopefully fixing those
> will
> > be
> > > >>> more aligned with the priorities of the community.
> > > >>>
> > > >>> @Katherin, my understanding is that given the limited resources,
> > there
> > > is
> > > >>> no development effort focused on batch processing right now.
> > > >>>
> > > >>> So to summarize, it seems like there are people willing to work on
> ML
> > > on
> > > >>> Flink, but nobody is sure how to do it.
> > > >>> There are many directions we could take (batch, online, batch on
> > > >>> streaming), each with its own merits and downsides.
> > > >>>
> > > >>> If you want we can start a design doc and move the conversation
> > there,
> > > >>> come
> > > >>> up with a roadmap and start implementing.
> > > >>>
> > > >>> Regards,
> > > >>> Theodore
> > > >>>
> > > >>> [1]
> > > >>> http://apache-flink-user-mailing-list-archive.2336050.n4.
> > > >>> nabble.com/Understanding-connected-streams-use-without-times
> > > >>> tamps-td10241.html
> > > >>>
> > > >>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <
> > [hidden email]
> > > >
> > > >>> wrote:
> > > >>>
> > > >>> It's great to see so much activity in this discussion :)
> > > >>>> I'll try to add my thoughts.
> > > >>>>
> > > >>>> I think building a developer community (Till's 2. point) can be
> > > slightly
> > > >>>> separated from what features we should aim for (1. point) and
> > > showcasing
> > > >>>> (3. point). Thanks Till for bringing up the ideas for
> restructuring,
> > > I'm
> > > >>>> sure we'll find a way to make the development process more
> dynamic.
> > > I'll
> > > >>>> try to address the rest here.
> > > >>>>
> > > >>>> It's hard to choose directions between streaming and batch ML. As
> > Theo
> > > >>>> has
> > > >>>> indicated, not much online ML is used in production, but Flink
> > > >>>> concentrates
> > > >>>> on streaming, so online ML would be a better fit for Flink.
> However,
> > > as
> > > >>>> most of you argued, there's definite need for batch ML. But batch
> ML
> > > >>>> seems
> > > >>>> hard to achieve because there are blocking issues with persisting,
> > > >>>> iteration paths etc. So it's no good either way.
> > > >>>>
> > > >>>> I propose a seemingly crazy solution: what if we developed batch
> > > >>>> algorithms also with the streaming API? The batch API would
> clearly
> > > seem
> > > >>>> more suitable for ML algorithms, but there a lot of benefits of
> this
> > > >>>> approach too, so it's clearly worth considering. Flink also has
> the
> > > high
> > > >>>> level vision of "streaming for everything" that would clearly fit
> > this
> > > >>>> case. What do you all think about this? Do you think this solution
> > > would
> > > >>>> be
> > > >>>> feasible? I would be happy to make a more elaborate proposal, but
> I
> > > push
> > > >>>> my
> > > >>>> main ideas here:
> > > >>>>
> > > >>>> 1) Simplifying by using one system
> > > >>>> It could simplify the work of both the users and the developers.
> One
> > > >>>> could
> > > >>>> execute training once, or could execute it periodically e.g. by
> > using
> > > >>>> windows. Low-latency serving and training could be done in the
> same
> > > >>>> system.
> > > >>>> We could implement incremental algorithms, without any side inputs
> > for
> > > >>>> combining online learning (or predictions) with batch learning. Of
> > > >>>> course,
> > > >>>> all the logic describing these must be somehow implemented (e.g.
> > > >>>> synchronizing predictions with training), but it should be easier
> to
> > > do
> > > >>>> so
> > > >>>> in one system, than by combining e.g. the batch and streaming API.
> > > >>>>
> > > >>>> 2) Batch ML with the streaming API is not harder
> > > >>>> Despite these benefits, it could seem harder to implement batch ML
> > > with
> > > >>>> the streaming API, but in my opinion it's not. There are more
> > > flexible,
> > > >>>> lower-level optimization potentials with the streaming API. Most
> > > >>>> distributed ML algorithms use a lower-level model than the batch
> API
> > > >>>> anyway, so sometimes it feels like forcing the algorithm logic
> into
> > > the
> > > >>>> training API and tweaking it. Although we could not use the batch
> > > >>>> primitives like join, we would have the E.g. in my experience with
> > > >>>> implementing a distributed matrix factorization algorithm [1], I
> > > couldn't
> > > >>>> do a simple optimization because of the limitations of the
> iteration
> > > API
> > > >>>> [2]. Even if we pushed all the development effort to make the
> batch
> > > API
> > > >>>> more suitable for ML there would be things we couldn't do. E.g.
> > there
> > > are
> > > >>>> approaches for updating a model iteratively without locks [3,4]
> > (i.e.
> > > >>>> somewhat asynchronously), and I don't see a clear way to implement
> > > such
> > > >>>> algorithms with the batch API.
> > > >>>>
> > > >>>> 3) Streaming community (users and devs) benefit
> > > >>>> The Flink streaming community in general would also benefit from
> > this
> > > >>>> direction. There are many features needed in the streaming API for
> > ML
> > > to
> > > >>>> work, but this is also true for the batch API. One really
> important
> > is
> > > >>>> the
> > > >>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot
> > of
> > > >>>> effort (mostly from Paris) for making it mature enough [6]. Kate
> > > >>>> mentioned
> > > >>>> using GPUs, and I'm sure they have uses in streaming generally
> [7].
> > > Thus,
> > > >>>> by improving the streaming API to allow ML algorithms, the
> streaming
> > > API
> > > >>>> benefit too (which is important as they have a lot more production
> > > users
> > > >>>> than the batch API).
> > > >>>>
> > > >>>> 4) Performance can be at least as good
> > > >>>> I believe the same performance could be achieved with the
> streaming
> > > API
> > > >>>> as
> > > >>>> with the batch API. Streaming API is much closer to the runtime
> than
> > > the
> > > >>>> batch API. For corner-cases, with runtime-layer optimizations of
> > batch
> > > >>>> API,
> > > >>>> we could find a way to do the same (or similar) optimization for
> the
> > > >>>> streaming API (see my previous point). Such case could be using
> > > managed
> > > >>>> memory (and spilling to disk). There are also benefits by default,
> > > e.g.
> > > >>>> we
> > > >>>> would have a finer grained fault tolerance with the streaming API.
> > > >>>>
> > > >>>> 5) We could keep batch ML API
> > > >>>> For the shorter term, we should not throw away all the algorithms
> > > >>>> implemented with the batch API. By pushing forward the development
> > > with
> > > >>>> side inputs we could make them usable with streaming API. Then, if
> > the
> > > >>>> library gains some popularity, we could replace the algorithms in
> > the
> > > >>>> batch
> > > >>>> API with streaming ones, to avoid the performance costs of e.g.
> not
> > > being
> > > >>>> able to persist.
> > > >>>>
> > > >>>> 6) General tools for implementing ML algorithms
> > > >>>> Besides implementing algorithms one by one, we could give more
> > general
> > > >>>> tools for making it easier to implement algorithms. E.g. parameter
> > > server
> > > >>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
> > > >>>> similar
> > > >>>> model to Flink streaming, we could look into that too. I think
> often
> > > when
> > > >>>> deploying a production ML system, much more configuration and
> > tweaking
> > > >>>> should be done than e.g. Spark MLlib allows. Why not allow that?
> > > >>>>
> > > >>>> 7) Showcasing
> > > >>>> Showcasing this could be easier. We could say that we're doing
> batch
> > > ML
> > > >>>> with a streaming API. That's interesting in its own. IMHO this
> > > >>>> integration
> > > >>>> is also a more approachable way towards end-to-end ML.
> > > >>>>
> > > >>>>
> > > >>>> Thanks for reading so far :)
> > > >>>>
> > > >>>> [1] https://github.com/apache/flink/pull/2819
> > > >>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
> > > >>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
> > > >>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
> > > >>>> 13-final77.pdf
> > > >>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
> > > >>>> Scoped+Loops+and+Job+Termination
> > > >>>> [6] https://github.com/apache/flink/pull/1668
> > > >>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.
> pdf
> > > >>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
> > > >>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
> > > >>>> com/Using-QueryableState-inside-Flink-jobs-and-
> > > >>>> Parameter-Server-implementation-td15880.html
> > > >>>>
> > > >>>> Cheers,
> > > >>>> Gabor
> > > >>>>
> > > >>>>
> > >
> > > --
> >
> > *Yours faithfully, *
> >
> > *Kate Eri.*
> >
>

[Gábor Hermann]

Okay, I've created a skeleton of the design doc for choosing a direction:
https://docs.google.com/document/d/1afQbvZBTV15qF3vobVWUjxQc49h3Ud06MIRhahtJ6dw/edit?usp=sharing

Much of the pros/cons have already been discussed here, so I'll try to
put there all the arguments mentioned in this thread. Feel free to put
there more :)

@Stavros: I agree we should take action fast. What about collecting our
thoughts in the doc by around Tuesday next week (28. February)? Then
decide on the direction and design a roadmap by around Friday (3.
March)? Is that feasible, or should it take more time?

I think it will be necessary to have a shepherd, or even better a
committer, to be involved in at least reviewing and accepting the
roadmap. It would be best, if a committer coordinated all this.
@Theodore: Would you like to do the coordination?

Regarding the use-cases: I've seen some abstracts of talks at SF Flink
Forward [1] that seem promising. There are companies already using Flink
for ML [2,3,4,5].

[1] http://sf.flink-forward.org/program/sessions/
[2]
http://sf.flink-forward.org/kb_sessions/experiences-with-streaming-vs-micro-batch-for-online-learning/
[3] http://sf.flink-forward.org/kb_sessions/introducing-flink-tensorflow/
[4]
http://sf.flink-forward.org/kb_sessions/non-flink-machine-learning-on-flink/
[5]
http://sf.flink-forward.org/kb_sessions/streaming-deep-learning-scenarios-with-flink/

Cheers,
Gabor


On 2017-02-23 15:19, Katherin Eri wrote:

> I have asked already some teams for useful cases, but all of them need time
> to think.
> During analysis something will finally arise.
> May be we can ask partners of Flink  for cases? Data Artisans got results
> of customers survey: [1], ML better support is wanted, so we could ask what
> exactly is necessary.
>
> [1] http://data-artisans.com/flink-user-survey-2016-part-2/
>
> 23 февр. 2017 г. 4:32 PM пользователь "Stavros Kontopoulos" <
> [hidden email]> написал:
>
>> +100 for a design doc.
>>
>> Could we also set a roadmap after some time-boxed investigation captured in
>> that document? We need action.
>>
>> Looking forward to work on this (whatever that might be) ;) Also are there
>> any data supporting one direction or the other from a customer perspective?
>> It would help to make more informed decisions.
>>
>> On Thu, Feb 23, 2017 at 2:23 PM, Katherin Eri <[hidden email]>
>> wrote:
>>
>>> Yes, ok.
>>> let's start some design document, and write down there already mentioned
>>> ideas about: parameter server, about clipper and others. Would be nice if
>>> we will also map this approaches to cases.
>>> Will work on it collaboratively on each topic, may be finally we will
>> form
>>> some picture, that could be agreed with committers.
>>> @Gabor, could you please start such shared doc, as you have already
>> several
>>> ideas proposed?
>>>
>>> чт, 23 февр. 2017, 15:06 Gábor Hermann <[hidden email]>:
>>>
>>>> I agree, that it's better to go in one direction first, but I think
>>>> online and offline with streaming API can go somewhat parallel later.
>> We
>>>> could set a short-term goal, concentrate initially on one direction,
>> and
>>>> showcase that direction (e.g. in a blogpost). But first, we should list
>>>> the pros/cons in a design doc as a minimum. Then make a decision what
>>>> direction to go. Would that be feasible?
>>>>
>>>> On 2017-02-23 12:34, Katherin Eri wrote:
>>>>
>>>>> I'm not sure that this is feasible, doing all at the same time could
>>> mean
>>>>> doing nothing((((
>>>>> I'm just afraid, that words: we will work on streaming not on
>> batching,
>>>> we
>>>>> have no commiter's time for this, mean that yes, we started work on
>>>>> FLINK-1730, but nobody will commit this work in the end, as it
>> already
>>>> was
>>>>> with this ticket.
>>>>>
>>>>> 23 февр. 2017 г. 14:26 пользователь "Gábor Hermann" <
>>>> [hidden email]>
>>>>> написал:
>>>>>
>>>>>> @Theodore: Great to hear you think the "batch on streaming" approach
>>> is
>>>>>> possible! Of course, we need to pay attention all the pitfalls
>> there,
>>>> if we
>>>>>> go that way.
>>>>>>
>>>>>> +1 for a design doc!
>>>>>>
>>>>>> I would add that it's possible to make efforts in all the three
>>>> directions
>>>>>> (i.e. batch, online, batch on streaming) at the same time. Although,
>>> it
>>>>>> might be worth to concentrate on one. E.g. it would not be so useful
>>> to
>>>>>> have the same batch algorithms with both the batch API and streaming
>>>> API.
>>>>>> We can decide later.
>>>>>>
>>>>>> The design doc could be partitioned to these 3 directions, and we
>> can
>>>>>> collect there the pros/cons too. What do you think?
>>>>>>
>>>>>> Cheers,
>>>>>> Gabor
>>>>>>
>>>>>>
>>>>>> On 2017-02-23 12:13, Theodore Vasiloudis wrote:
>>>>>>
>>>>>>> Hello all,
>>>>>>>
>>>>>>>
>>>>>>> @Gabor, we have discussed the idea of using the streaming API to
>>> write
>>>> all
>>>>>>> of our ML algorithms with a couple of people offline,
>>>>>>> and I think it might be possible and is generally worth a shot. The
>>>>>>> approach we would take would be close to Vowpal Wabbit, not exactly
>>>>>>> "online", but rather "fast-batch".
>>>>>>>
>>>>>>> There will be problems popping up again, even for very simple algos
>>>> like
>>>>>>> on
>>>>>>> line linear regression with SGD [1], but hopefully fixing those
>> will
>>> be
>>>>>>> more aligned with the priorities of the community.
>>>>>>>
>>>>>>> @Katherin, my understanding is that given the limited resources,
>>> there
>>>> is
>>>>>>> no development effort focused on batch processing right now.
>>>>>>>
>>>>>>> So to summarize, it seems like there are people willing to work on
>> ML
>>>> on
>>>>>>> Flink, but nobody is sure how to do it.
>>>>>>> There are many directions we could take (batch, online, batch on
>>>>>>> streaming), each with its own merits and downsides.
>>>>>>>
>>>>>>> If you want we can start a design doc and move the conversation
>>> there,
>>>>>>> come
>>>>>>> up with a roadmap and start implementing.
>>>>>>>
>>>>>>> Regards,
>>>>>>> Theodore
>>>>>>>
>>>>>>> [1]
>>>>>>> http://apache-flink-user-mailing-list-archive.2336050.n4.
>>>>>>> nabble.com/Understanding-connected-streams-use-without-times
>>>>>>> tamps-td10241.html
>>>>>>>
>>>>>>> On Tue, Feb 21, 2017 at 11:17 PM, Gábor Hermann <
>>> [hidden email]
>>>>>>> wrote:
>>>>>>>
>>>>>>> It's great to see so much activity in this discussion :)
>>>>>>>> I'll try to add my thoughts.
>>>>>>>>
>>>>>>>> I think building a developer community (Till's 2. point) can be
>>>> slightly
>>>>>>>> separated from what features we should aim for (1. point) and
>>>> showcasing
>>>>>>>> (3. point). Thanks Till for bringing up the ideas for
>> restructuring,
>>>> I'm
>>>>>>>> sure we'll find a way to make the development process more
>> dynamic.
>>>> I'll
>>>>>>>> try to address the rest here.
>>>>>>>>
>>>>>>>> It's hard to choose directions between streaming and batch ML. As
>>> Theo
>>>>>>>> has
>>>>>>>> indicated, not much online ML is used in production, but Flink
>>>>>>>> concentrates
>>>>>>>> on streaming, so online ML would be a better fit for Flink.
>> However,
>>>> as
>>>>>>>> most of you argued, there's definite need for batch ML. But batch
>> ML
>>>>>>>> seems
>>>>>>>> hard to achieve because there are blocking issues with persisting,
>>>>>>>> iteration paths etc. So it's no good either way.
>>>>>>>>
>>>>>>>> I propose a seemingly crazy solution: what if we developed batch
>>>>>>>> algorithms also with the streaming API? The batch API would
>> clearly
>>>> seem
>>>>>>>> more suitable for ML algorithms, but there a lot of benefits of
>> this
>>>>>>>> approach too, so it's clearly worth considering. Flink also has
>> the
>>>> high
>>>>>>>> level vision of "streaming for everything" that would clearly fit
>>> this
>>>>>>>> case. What do you all think about this? Do you think this solution
>>>> would
>>>>>>>> be
>>>>>>>> feasible? I would be happy to make a more elaborate proposal, but
>> I
>>>> push
>>>>>>>> my
>>>>>>>> main ideas here:
>>>>>>>>
>>>>>>>> 1) Simplifying by using one system
>>>>>>>> It could simplify the work of both the users and the developers.
>> One
>>>>>>>> could
>>>>>>>> execute training once, or could execute it periodically e.g. by
>>> using
>>>>>>>> windows. Low-latency serving and training could be done in the
>> same
>>>>>>>> system.
>>>>>>>> We could implement incremental algorithms, without any side inputs
>>> for
>>>>>>>> combining online learning (or predictions) with batch learning. Of
>>>>>>>> course,
>>>>>>>> all the logic describing these must be somehow implemented (e.g.
>>>>>>>> synchronizing predictions with training), but it should be easier
>> to
>>>> do
>>>>>>>> so
>>>>>>>> in one system, than by combining e.g. the batch and streaming API.
>>>>>>>>
>>>>>>>> 2) Batch ML with the streaming API is not harder
>>>>>>>> Despite these benefits, it could seem harder to implement batch ML
>>>> with
>>>>>>>> the streaming API, but in my opinion it's not. There are more
>>>> flexible,
>>>>>>>> lower-level optimization potentials with the streaming API. Most
>>>>>>>> distributed ML algorithms use a lower-level model than the batch
>> API
>>>>>>>> anyway, so sometimes it feels like forcing the algorithm logic
>> into
>>>> the
>>>>>>>> training API and tweaking it. Although we could not use the batch
>>>>>>>> primitives like join, we would have the E.g. in my experience with
>>>>>>>> implementing a distributed matrix factorization algorithm [1], I
>>>> couldn't
>>>>>>>> do a simple optimization because of the limitations of the
>> iteration
>>>> API
>>>>>>>> [2]. Even if we pushed all the development effort to make the
>> batch
>>>> API
>>>>>>>> more suitable for ML there would be things we couldn't do. E.g.
>>> there
>>>> are
>>>>>>>> approaches for updating a model iteratively without locks [3,4]
>>> (i.e.
>>>>>>>> somewhat asynchronously), and I don't see a clear way to implement
>>>> such
>>>>>>>> algorithms with the batch API.
>>>>>>>>
>>>>>>>> 3) Streaming community (users and devs) benefit
>>>>>>>> The Flink streaming community in general would also benefit from
>>> this
>>>>>>>> direction. There are many features needed in the streaming API for
>>> ML
>>>> to
>>>>>>>> work, but this is also true for the batch API. One really
>> important
>>> is
>>>>>>>> the
>>>>>>>> loops API (a.k.a. iterative DataStreams) [5]. There has been a lot
>>> of
>>>>>>>> effort (mostly from Paris) for making it mature enough [6]. Kate
>>>>>>>> mentioned
>>>>>>>> using GPUs, and I'm sure they have uses in streaming generally
>> [7].
>>>> Thus,
>>>>>>>> by improving the streaming API to allow ML algorithms, the
>> streaming
>>>> API
>>>>>>>> benefit too (which is important as they have a lot more production
>>>> users
>>>>>>>> than the batch API).
>>>>>>>>
>>>>>>>> 4) Performance can be at least as good
>>>>>>>> I believe the same performance could be achieved with the
>> streaming
>>>> API
>>>>>>>> as
>>>>>>>> with the batch API. Streaming API is much closer to the runtime
>> than
>>>> the
>>>>>>>> batch API. For corner-cases, with runtime-layer optimizations of
>>> batch
>>>>>>>> API,
>>>>>>>> we could find a way to do the same (or similar) optimization for
>> the
>>>>>>>> streaming API (see my previous point). Such case could be using
>>>> managed
>>>>>>>> memory (and spilling to disk). There are also benefits by default,
>>>> e.g.
>>>>>>>> we
>>>>>>>> would have a finer grained fault tolerance with the streaming API.
>>>>>>>>
>>>>>>>> 5) We could keep batch ML API
>>>>>>>> For the shorter term, we should not throw away all the algorithms
>>>>>>>> implemented with the batch API. By pushing forward the development
>>>> with
>>>>>>>> side inputs we could make them usable with streaming API. Then, if
>>> the
>>>>>>>> library gains some popularity, we could replace the algorithms in
>>> the
>>>>>>>> batch
>>>>>>>> API with streaming ones, to avoid the performance costs of e.g.
>> not
>>>> being
>>>>>>>> able to persist.
>>>>>>>>
>>>>>>>> 6) General tools for implementing ML algorithms
>>>>>>>> Besides implementing algorithms one by one, we could give more
>>> general
>>>>>>>> tools for making it easier to implement algorithms. E.g. parameter
>>>> server
>>>>>>>> [8,9]. Theo also mentioned in another thread that TensorFlow has a
>>>>>>>> similar
>>>>>>>> model to Flink streaming, we could look into that too. I think
>> often
>>>> when
>>>>>>>> deploying a production ML system, much more configuration and
>>> tweaking
>>>>>>>> should be done than e.g. Spark MLlib allows. Why not allow that?
>>>>>>>>
>>>>>>>> 7) Showcasing
>>>>>>>> Showcasing this could be easier. We could say that we're doing
>> batch
>>>> ML
>>>>>>>> with a streaming API. That's interesting in its own. IMHO this
>>>>>>>> integration
>>>>>>>> is also a more approachable way towards end-to-end ML.
>>>>>>>>
>>>>>>>>
>>>>>>>> Thanks for reading so far :)
>>>>>>>>
>>>>>>>> [1] https://github.com/apache/flink/pull/2819
>>>>>>>> [2] https://issues.apache.org/jira/browse/FLINK-2396
>>>>>>>> [3] https://people.eecs.berkeley.edu/~brecht/papers/hogwildTR.pdf
>>>>>>>> [4] https://www.usenix.org/system/files/conference/hotos13/hotos
>>>>>>>> 13-final77.pdf
>>>>>>>> [5] https://cwiki.apache.org/confluence/display/FLINK/FLIP-15+
>>>>>>>> Scoped+Loops+and+Job+Termination
>>>>>>>> [6] https://github.com/apache/flink/pull/1668
>>>>>>>> [7] http://lsds.doc.ic.ac.uk/sites/default/files/saber-sigmod16.
>> pdf
>>>>>>>> [8] https://www.cs.cmu.edu/~muli/file/parameter_server_osdi14.pdf
>>>>>>>> [9] http://apache-flink-mailing-list-archive.1008284.n3.nabble.
>>>>>>>> com/Using-QueryableState-inside-Flink-jobs-and-
>>>>>>>> Parameter-Server-implementation-td15880.html
>>>>>>>>
>>>>>>>> Cheers,
>>>>>>>> Gabor
>>>>>>>>
>>>>>>>>
>>>> --
>>> *Yours faithfully, *
>>>
>>> *Kate Eri.*
>>>