Rhea: Automatic Filtering for Unstructured Cloud Storage
summary
Rhea performs static analysis of Hadoop Java bytecode to extract row filters and column filters as executable, conservative, stateless predicates that run on storage nodes (Java methods and optimized C for columns). Filters process ~100 MB/s per core (optimized column code 5–70× faster) and typically need ~2 cores (one job needs ~10) to avoid being the bottleneck. Measured effects: 2x–20,000x reduction in data transfer, up to 5× faster job runs and up to 13× lower cross‑cloud egress cost; filters are best‑effort, safe, and transparent.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 0:18Traditional Data Analytics
- 1:44Job Interest
- 2:23Summary
- 2:51Strongman approach
- 3:59The problem
- 4:25Example
- 5:38Java
- 6:27Overview
- 7:46Raw Filters
- 9:28Column Filters
- 10:19Experiments
- 12:49Experiment
- 15:15Runtime
- 15:58Recap
- 16:47QA
