DBTB INT Chris Fregly
summary
Argues that for high-throughput streaming workloads exact answers cost too much and probabilistic approximations often meet production needs, demonstrates this with Spark and TensorFlow examples, digs into Project Tungsten internals down to the JVM and C/OS layer to show where bottlenecks occur, and advocates exploring dataset structure before modelling and making domain-specific tweaks to models rather than using off‑the‑shelf defaults; presents these changes as concrete fixes that reduce latency and resource use in practice.
AI-written summary · about
related talks
47mExpert Panel: Current & Future State of the Art in Real-Time Data Infrastructure
32mBig Data Processing : Where do we go next ?
12mGetting started with fast data architecture
34mWorking with Tensorflow from the JVM #bbuzz
10mRoad to Real-Time Digital Business
29mDataEngBytes 2023 - BRI-T1-03 - Yousry Mohamed
06mInterview with Ted Dunning
45mData Stream Processing Concepts and Implementations