ML Inference at scale, with Apache Beam and the Flink runner: Easy as learning your 5 times tables.
summary
Builds beam pipelines that run TensorFlow models with the Flink runner using the RunInference transform from tfx bsl, loading SavedModel from GCS in local mode and on an embedded Flink cluster. Shows converting inputs to serialized tf.Example via a tf.function signature so RunInference accepts BigQuery rows and TFRecord sources, demonstrates key forwarding and batching, runs two models in parallel and as a cascade ensemble, and prints correct numeric outputs (five‑times and ten‑times results) end to end.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:29Overview
- 1:24ML tasks
- 2:50Execution environments
- 3:53Portability
- 7:26Apache Beam
- 11:41RunInference
- 24:32Data Warehouse
- 27:49Live Experimentation
- 31:11Cascade Ensembles
- 34:49Flink Runner
- 36:16Authoring components
related talks
36mMachine Learning with Apache Beam
40mStreaming ML with Flink
37mBuild a Flink AI Ecosystem
38mDistributed Processing for Machine Learning Production Pipelines
21mIntroducing FlinkML
37mSimplifying Training Deep & Serving Learning Models...using Tensorflow
19mBeam Up Your GenAI Usage: Usability, Efficiency, Reliability
21mPredicting Hacker News with Beam and TensorFlow