Flink-JPMML: An open source Flink library for Streaming ML Model Serving
summary
An open source Flink operator that serves PMML models for streaming predictions, loading models on demand from HDFS or a model repo via a broadcast control stream and keeping a per-slot metadata table for fault tolerant state. Events must carry a model ID, the operator converts inputs to vector form, runs JPMML evaluation, normalizes outputs and reports per-version accuracy (demo shows ~40–80% and a newer version ≈70%). Limitations: no online learning, numeric-only polymorphism issues, but the live demo for multi-version comparison works.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:17What is Model Serving
- 1:49Why PMML
- 3:10Architecture
- 4:22FlinkJPMML
- 6:11Dynamic behavior
- 7:56Contra stream
- 9:59Control stream
- 10:50Metadata table
- 11:54Events
- 14:14Checkpointing
- 17:06Whats next
- 18:30Testing
- 21:05Creating Streams
- 22:20Model Pot
- 23:25Conscious Team
- 25:55Output to normalization
related talks
40mStreaming ML with Flink
15mMetamodelization and Automatic Generation of ML Pipelines
28mAMIDST Toolbox: Scalable probabil. machine learning w/ Flink
15mData science at scale using Apache Flink: Dynamic model serving and real-time feature generation
24mMLflow: An open platform to simplify...
31mMLflow: Platform for Complete Machine Learning Lifecycle
39mML Inference at scale, with Apache Beam and the Flink runner: Easy as learning your 5 times tables.
41mOpen standards for machine learning deployment