Project for CSE 584 - Advanced Database Systems (Winter 2025) at the University of Michigan under Prof. Lin Ma. Done in collaboration with Anirudh Bhaskar, Divyam Sharma, and Nilay Gautam.

Delta Lake provides ACID guarantees and efficient metadata management for single-table workloads on cloud object stores, but lacks native support for atomic operations spanning multiple tables. In this work, we introduce transaction groups (or T-Groups), a protocol extension that enables multi-table transactions using only object-store primitives — no external coordinators required. We implement our design in delta-rs-mt, a fork of the delta-rs library, and evaluate on TPC-DS benchmarks. We find that single-table operations incur <0.25s of write overhead and reads show <3% latency increase; multi-table writes exhibit predictable log-contention. Our approach seamlessly integrates with existing Delta Lake features, making robust multi-table ACID transactions practical for large-scale analytics without relying on external coordination services.

The code is available on github and the full report is available below.