Master large-scale ML batch inferencing
summary
Nearmap runs a pipeline-of-pipelines on 10,000-node EKS cluster using Kubeflow, Kafka, RDS and FastAPI to process imagery at scale, storing 100 PB in S3 and producing ~150 semantic layers per tile. Data immutability uses Merkel tree hashing to guarantee data-level idempotency; workloads target spot instances and scale from 1–5M tiles/day to spikes of ~30M. An emergency run ingests 118 flights, extracts 5.3M buildings, and delivers results within ~3–4 hours.
AI-written summary · about
chapters (8)
- 0:00Introduction and welcome
- 1:18MLOps vs DevOps foundations
- 6:40Batch inferencing strategies
- 12:20Near Maps overview and scale
- 19:18Data level item potency
- 21:05Scaling to 10,000 nodes
- 26:23Real-world impact and case study
- 29:11Q&A and closing
related talks
29mRunning Massively Parallel Deep-learning Inference Pipelines on Kub... Suneeta Mall & Martin Abeleda
30mRealizing End to End Reproducible Machine Learning on Kubernetes
44mML in Production
28mScaling Machine Learning Training and Inference
46mAccelerating Machine Learning DevOps with Kubeflow
34mDeep Learning at Scale
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow