Streaming ML with Flink
summary
FlinkML gets a Scala extension that provides streaming predictors so a batch-trained model (P and Q factor matrices from ALS) is serialized (prototype write model to HDFS) and loaded into a stateful streaming operator to serve predictions in a few hundred milliseconds. Implementation adds implicit conversions and a predict(operator) API; several prototypes exist (implicit-ALS tested, matrix factorization prototype, logistic regression partly done needing SGD). Experiments on a public dataset show online SGD outperforms batch under concept drift, so streaming training is promising though still research.
AI-written summary · about
chapters (7)
- 0:00Introduction to streaming ML
- 2:57Flink ML library overview
- 7:30Matrix factorization theory
- 11:49Implementing streaming predictors
- 19:41Online learning research
- 22:13Streamline project use cases
- 30:35Summary and Q&A
related talks
36mFlink-JPMML: An open source Flink library for Streaming ML Model Serving
21mIntroducing FlinkML
31mExtending Apache Flink stream processing with Apache Samoa ML methods
28mEnabling Machine Learning with Apache Flink
37mBuild a Flink AI Ecosystem
24mUnder Pressure: Applying ML in Real Time
39mML Inference at scale, with Apache Beam and the Flink runner: Easy as learning your 5 times tables.
31mLeveraging GraphQL for Continual Learning in Real-Time ML Systems