Seldon – Serves predictions via a REST API | Hacker News Reader

Seldon – Serves predictions via a REST API

github.com

70 pointspigloo26 comments

Hey folks, thanks for your interest in Seldon. It's exciting to see all of our hard work come to life!

Our plan is to build out and support the strongest ecosystem of data scientists in the predictive machine learning space. We adopted the permissive Apache 2.0 license, so you are free to build domain-specific algorithms and functionality to either contribute back into the core platform or release under a separate license.

We would love to hear your feedback, both first impressions and thoughts as you start using the platform. And, of course, how you would like to see the project evolve in the long term.

Suggestions on the functionality that you would find useful would also be appreciated - we're hooked on customer and developer community feedback! :)

Alex, The idea is very cool and I am glad that the developments are going to be contributed to an open source version.

As a developer while we can try VM setup and as well as Amazon instance, can you also host it on one of your server and provide a demo access to developers to get acquainted with system and architecture before they can try on VM or Amazon AMI?

Thanks, Hassan!

We were originally a SaaS recommendation engine and will continue to provide a fully-managed cloud-based solution. It’s now serving recommendations on hundreds of millions of page views per month.

There are a few pretty significant technical tasks we need to complete to make the API provisioning streamlined enough to support trial accounts since there are some manual steps required to configure the metadata properly, etc.

Since we decided to go open-source last September, all of our R&D focus has been in preparing this open-source release instead of the self-serve. Streamlining our cloud service is in the roadmap, and it's now powered by the same codebase as the open-source project.

The feedback we received from developers who tried the AMI and VM are that it's super quick to get started. And as you can see we've spent a lot of time recently on documentation.

If you're interested in the hosted API, please email support@seldon.io.

Will this project cope with the appearance of The Mule?

Perhaps if you run it on FoundationDB, it can fall back to a second FoundationDB server in the event of a crisis.

Looks interesting, seems like a neat product.

Some thoughts on how to make the webpage more informative:

- Improve the schematic overview[1], e.g. what's the purpose of a logging cluster? General logging? The MySQL database, what's it's role? My initial reaction is that logging/sql-storage could be handled outside of Seldon. I may be wrong, but you should help people here.

- Same with load-balancers, seems like it should be outside the scope of a prediction service. If not, an explanation would help.

- Clarify which parts are open-source, and which ones are closed. E.g. with ElasticSearch, it's obvious on their front-page which components are open-source and which-ones are paid.

- Add a comparison to prediction.io (RethinkDB's comparison with MongoDB[2] is a good example). It's an obvious question people will have, answering it upfront is helpful. Keep it fair/unbiased.

[1] http://www.seldon.io/open-source/ [2] http://rethinkdb.com/docs/comparison-tables/

Thanks - this is all useful feedback.

You're right about the load balancers - this doesn't form part of the project codebase, but it's important for people to know that the API servers are stateless and can scale horizontally. However, the other infrastructure components are required as Seldon provides the full end-to-end platform not just collection of predictive algorithms.

We will review your comments and for now we linked to the infrastructure page in our docs: http://docs.seldon.io/tech.html

We are planning to release some functionality under a separate commercial license, just as anyone is free to do. Everything you currently see on the website and Github repos is open-source and licensed under Apache 2.0.

This seems interesting. Does it support user-defined algorithms? Or does it have to be a combination of the exsiting algorithms that Seldon provides? I don't seem to able to get a definite answer from this page: http://docs.seldon.io/configuration.html

We are targeting to provide this in the next few weeks. Sign up for the newsletter or follow the project and we'll notify you when it's available.

Hi, I'm a developer on this project so I can answer any technical questions you may have.

Explain to me like you are talking to a 5 year old. What can I do with this?

He offered to answer technical questions, so let me step in on this one: it's a recommendation engine. E.g., it can do things like Amazon's "you may also like" feature.

Try this:

http://docs.seldon.io/index.html

I've not heard of Seldon before. I'd love to hear your 'elevator pitch' for it?

Please check:

http://docs.seldon.io/

http://www.seldon.io/algorithms/

Seldon is an open-source predictive machine learning platform that includes a high-performance recommendation engine and data enrichment. It can run multiple algorithms and configurations to optimise KPIs. It will shortly include a pluggable architecture to allow data scientists to deploy their custom algorithms.

Hey this looks sweet. I'm still new to this field, but I'm looking to do some collaborative filtering across large datasets (4 billion rows+)

I was looking at using Prediction.IO and this looks suitable too, can you elaborate on the high level differences between this project and prediction.io?

I'm also curious about horizontal scaling.

The main reason why Seldon is different from other open-source prediction engines: * Seldon have come from a background of high-scale enterprise deployments before releasing an open source platform, not the other way around. In other words, we already optimized for real-time low latency high throughput production environments. * Seldon provides an end-to-end componentized setup - i.e. front end UI, real-time prediction server, offline machine learning jobs, web-scraping, etc. * Seldon allows developers to run A/B tests and change algorithms with no downtime. * Seldon will provide enterprise features like automated performance optimization of customer's own algorithms using micro-services.

Presently most of our CF algorithms utilize Apache Spark but we intend to be agnostic on this and allow any machine learning platform to be integrated. I believe that Spark can easily handle this size of data set.

With regards to horizontal scaling, there are two parts to consider. Creating the models and serving the recommendations (the Seldon server project).

Model creation is done in a variety of ways, but but can be managed with scalable technologies such as Spark.

The Seldon Server project can be deployed on as many machines are you require and they will work together to provide recommendations behind a load balancer. We have experience working with some very large news websites so this part of our technology is well developed.

Great thanks for your feedback, I will setup an installation in my home lab this weekend, super excited!

I maintain a small search engine that allows users to search for any file in my organization's servers. I'm primarily looking for a tool that will help me model user's search queries and user characteristics. Specifically clustering their queries into groups like searching for Documents, movies, or music, as well as what the client's OS is. Can Seldon do this easily for me, or is not the right program for the job, or even overkill?

It sounds like you basically want a form of document clustering. Seldon has integrated the Semantic Vectors https://code.google.com/p/semanticvectors/ project which can be used to do this or depending on your technical level to look directly at Semantic Vectors or toolkits such as gensim https://radimrehurek.com/gensim/

Great name - nice Asimov reference

How does this compare to PredictionIO?

It looks like they have built an interesting platform, but to be honest we haven't really paid attention to the details. We've been 100% focussed on our own product, but here’s my reply to a similar question which covers some of the unique benefits of Seldon: https://news.ycombinator.com/item?id=9262971

I had no idea what this was, but found the main web site here: http://www.seldon.io/ (It features a carousel which updates faster than you can casually read the text, which I did not find encouraging.)

Thanks for highlighting this issue, we've slowed down the carousel.

Reply on news.ycombinator.com