Weekly Paper Notes — three entries per week: 1 seminal + 2 top picks.
🔁 Seminal Paper of the Week
- Access Path Selection in a Relational Database Management System: Pat Selinger, Morton Astrahan, Donald Chamberlin, Raymond Lorie, Thomas Price, IBM Research (SIGMOD, 1979). The System R optimizer: a page-fetch plus CPU cost formula, selectivity factors computed from catalog statistics, and dynamic programming over join orders that keeps the best plan for each “interesting order”. Postgres and Spark still plan joins this way.
⭐ Top Papers of the Week
- Unbundling Multimodal Vision Encoding: Brown, Hoang, …, Zettlemoyer, Beyer, Ren, Verbeek, Xie, Xu, Meta FAIR / NYU (arXiv:2611.17334). With every model trained from scratch at matched lifetime compute, a pretrained vision encoder adds nothing, but dedicated vision parameters do: encoder-free models need about 3× the compute to catch up. Vision-only branches in the first eight backbone layers match a separate encoder and cut time to first token.
- Raven: Towards Unified File Management for Data Lakehouse Systems: Oh, Mehra, Kimball, Deshpande, University of Maryland / Dolby (arXiv:2611.15975). Snapshot membership and references go in a database, per-file statistics stay in Iceberg or Delta. The split gives 55–61% faster small commits, multi-table snapshot-isolation transactions at 146 txn/s against DuckLake’s 25, and reachability-based GC that deletes about 2,200 files/s.
The full weekly Discord digest covered ~15 more papers across AI/ML, NLP, distributed computing, databases, OS and networking; only the highest-impact picks are expanded into blog notes.
