Thanks for the feedback. If we strip away all of the fancy stuff, my proposal boils down to exposing the logic used in Spark's ML library. In an ideal world, Spark would possibly have relied on an existing ML implementation rather than reimplement, since there's very little that's Spark specific about using ML models. As Sean says, it may make most sense to have localPipelines live outside of Spark. However it would be really beneficial for Spark ML pipelines adoption if they used non-Spark logic. This would eliminate issues with train-serve skew and close the potential for bugs.
I'll leave some more comments in-line to Sean's response: I'm skeptical. Serving synchronous queries from a model at scale is a fundamentally different activity. As you note, it doesn't logically involve Spark. If it has to happen in milliseconds it's going to be in-core. Scoring even 10qps with a Spark job per request is probably a non-starter; think of the thousands of tasks per second and the overhead of just tracking them. When you say the RDDs support point prediction, I think you mean that those older models expose a method to score a Vector. They are not somehow exposing distributed point prediction. You could add this to the newer models, but it raises the question of how to make the Row to feed it? the .mllib punts on this and assumes you can construct the Vector. AK: In my mind, punting is exactly the right solution - no overhead, full control to the user I think this sweeps a lot under the rug in assuming that there can just be a "local" version of every Transformer -- but, even if there could be, consider how much extra implementation that is. Lots of them probably could be but I'm not sure that all can. AK: I'm not aware of models for which this is not possible - there are no Spark-only algorithms that I'm aware of. The work to convert Spark to Local models may be more involved for some implementations, sure, but I don't think any would be too bad. However if there is something that's impossible, then that's fine too. I'm not sure we have to commit to having local versions for every single model The bigger problem in my experience is the Pipelines don't generally encapsulate the entire pipeline from source data to score. They encapsulate the part after computing underlying features. That is, if one of your features is "total clicks from this user", that's the product of a DataFrame operation that precedes a Pipeline. This can't be turned into a non-distributed, non-Spark local version. AK: That's a great point, and a really good argument for keeping any local pipeline logic outside of Spark Solving subsets of this problem could still be useful, and you've highlighted some external projects that try. I'd also highlight PMML as an established interchange format for just the model part, and for cases that don't involve much or any pipeline, it's a better fit paired with a library that can score from PMML. AK: The problem with solutions like PMML is that they can tell you WHAT to do, but not HOW EXACTLY to do it. At the end of the day, the best model-description possible would be the metadata+ the code itself. That's the crux of my proposal - expose the implementation so users can use Spark models with the same exact code that was used to train I think this is one of those things that could live outside the project, because it's more not-Spark than Spark. Remember too that building a solution into the project blesses one at the expense of others. Asher Krim Senior Software Engineer On Mon, Mar 13, 2017 at 11:08 AM, Dongjin Lee <dong...@apache.org> wrote: > Although I love the cool idea of Asher, I'd rather +1 for Sean's view; I > think it would be much better to live outside of the project. > > Best, > Dongjin > > On Mon, Mar 13, 2017 at 5:39 PM, Sean Owen <so...@cloudera.com> wrote: > >> I'm skeptical. Serving synchronous queries from a model at scale is a >> fundamentally different activity. As you note, it doesn't logically involve >> Spark. If it has to happen in milliseconds it's going to be in-core. >> Scoring even 10qps with a Spark job per request is probably a non-starter; >> think of the thousands of tasks per second and the overhead of just >> tracking them. >> >> When you say the RDDs support point prediction, I think you mean that >> those older models expose a method to score a Vector. They are not somehow >> exposing distributed point prediction. You could add this to the newer >> models, but it raises the question of how to make the Row to feed it? the >> .mllib punts on this and assumes you can construct the Vector. >> >> I think this sweeps a lot under the rug in assuming that there can just >> be a "local" version of every Transformer -- but, even if there could be, >> consider how much extra implementation that is. Lots of them probably could >> be but I'm not sure that all can. >> >> The bigger problem in my experience is the Pipelines don't generally >> encapsulate the entire pipeline from source data to score. They encapsulate >> the part after computing underlying features. That is, if one of your >> features is "total clicks from this user", that's the product of a >> DataFrame operation that precedes a Pipeline. This can't be turned into a >> non-distributed, non-Spark local version. >> >> Solving subsets of this problem could still be useful, and you've >> highlighted some external projects that try. I'd also highlight PMML as an >> established interchange format for just the model part, and for cases that >> don't involve much or any pipeline, it's a better fit paired with a library >> that can score from PMML. >> >> I think this is one of those things that could live outside the project, >> because it's more not-Spark than Spark. Remember too that building a >> solution into the project blesses one at the expense of others. >> >> >> On Sun, Mar 12, 2017 at 10:15 PM Asher Krim <ak...@hubspot.com> wrote: >> >>> Hi All, >>> >>> I spent a lot of time at Spark Summit East this year talking with Spark >>> developers and committers about challenges with productizing Spark. One of >>> the biggest shortcomings I've encountered in Spark ML pipelines is the lack >>> of a way to serve single requests with any reasonable performance. >>> SPARK-10413 explores adding methods for single item prediction, but I'd >>> like to explore a more holistic approach - a separate local api, with >>> models that support transformations without depending on Spark at all. >>> >>> I've written up a doc >>> <https://docs.google.com/document/d/1Ha4DRMio5A7LjPqiHUnwVzbaxbev6ys04myyz6nDgI4/edit?usp=sharing> >>> detailing the approach, and I'm happy to discuss alternatives. If this >>> gains traction, I can create a branch with a minimal example on a simple >>> transformer (probably something like CountVectorizerModel) so we have >>> something concrete to continue the discussion on. >>> >>> Thanks, >>> Asher Krim >>> Senior Software Engineer >>> >> > > > -- > *Dongjin Lee* > > > *Software developer in Line+.So interested in massive-scale machine > learning.facebook: www.facebook.com/dongjin.lee.kr > <http://www.facebook.com/dongjin.lee.kr>linkedin: > kr.linkedin.com/in/dongjinleekr > <http://kr.linkedin.com/in/dongjinleekr>github: > <http://goog_969573159/>github.com/dongjinleekr > <http://github.com/dongjinleekr>twitter: www.twitter.com/dongjinleekr > <http://www.twitter.com/dongjinleekr>* >