big data spain - nov 17 2016 - madrid continuously deploy spark ml and tensorflow ai models: from...

20
BIG DATA SPAIN 2016 Continuously Deploy ML and AI Models: From Notebook to Microservice Thank You, Madrid! Chris Fregly, Research Scientist @

Upload: chris-fregly

Post on 15-Jan-2017

782 views

Category:

Software


0 download

TRANSCRIPT

BIG DATA SPAIN 2016

Continuously Deploy ML and AI Models: From Notebook to Microservice

Thank You, Madrid!Chris Fregly, Research Scientist @

WHO AM I?

Chris Fregly--------

Research Scientist @ PipelineIO(Formerly Netflix and Databricks)

--------http://pipeline.io

WORKSHOP ON SATURDAY, NOV 19THHERE IN MADRID!!

http://pipeline.io

SOURCE CODE AND DOCKER IMAGES

• Github Repo: ~900 Stars, ~300 Forks

• DockerHub Repo: ~ 6,000 Pulls

WHAT IS PIPELINE.IO?

Extending Your ML Pipelines into Production

100% Open Source!

http://pipeline.io

BRAINSTORMING AND VALIDATING

• Major Gaming Company

• Large Ride Sharing Service

• Popular Q & A Site

• Online Clothing Retailer

• Dominant Video Streaming

PIPELINE.IO FOCUS

• Model Deploying and Testing

• Model Scaling and Serving

• Online Model Training

• Native Code Generation

MODEL DEPLOYING AND TESTING

Continuously Test and Deploy Models

in Production!

MODEL SCALING AND SERVING

ONLINE MODEL TRAINING

• Continuous, Incremental, and Partial Training

• Kafka + Spark Streaming + Spark ML

• Real-time, Dynamic Recommendations

NATIVE CODE GENERATION

Generate OptimizedCode from Spark ML!

BECOME A CONTRIBUTOR!

PIPELINE.IO FOCUS FOR 2017

• Performance, Performance, Performance

• Native Code Generation: CPU + GPU

• More Global Contributors!

WE’RE HIRING!!

• Kafka, Spark ML, and TensorFlow Contributors

• Systems Engineers

• GPU/CUDA Engineers

• C++, Java, Scala, Python

WE ONLY HIRENICE PEOPLE!!

DEMO TIME!

DEMO: NETFLIX-BASED MICROSERVICES

Circuit Breakers and Request Batching

DEMO: DEPLOY NOTEBOOK TO PROD

Deploy to Cloudor On-Premise!

DEMO: NATIVE CODE GENERATION

DEMO: TENSORFLOW SERVING

THANK YOU FOR CHOOSING ME!

http://pipeline.io