South Bay Systems: File and Storage Technologies
Welcome to another edition of South Bay Systems! This time we bring you three wonderful talks from authors at the just-finishing Conference on File and Storage Technologies (FAST '26):
Swapnil Pimpale will present his work on Apple's Object Store.
Xinjun Yang will present his work on PolarStore, the new compression in PolarDB.
Ilya Meignan--Masson will present his work uCache, a unikernel-based I/O cache.
Agenda
6:00 PM: Doors open, food and socializing
6:30 PM — 6:50 PM: Apple Object Store Talk
6:50 PM — 7:10 PM: PolarStore Talk
7:10 PM — 7:30 PM: uCache Talk
7:30 PM onward : Community socializing!
Food and beverages will be provided, courtesy of our hosts, PureStorage.
Apple’s Geo-Distributed Object Store at Exabyte Scale
Over the last two decades, with the advent of mobile computing and Internet streaming, Apple has expanded its user base and services significantly. With this growth, we have seen an increased volume and diversity of data storage, ranging from backups, personal photos and videos, to music libraries, TV shows, and live streaming. In this paper, we present Apple Cloud Object Storage (ACOS), Apple’s object store designed to meet the specific requirements of user-facing and internal services, by accommodating a wide range of content and access patterns. ACOS has been in production for over a decade, storing several exabytes of objects and serving billions of requests per day. With its geo-replicated architecture using both local and regional replication mechanisms, ACOS is cost-efficient and highly scalable, available, and durable. The evaluation of our production deployment shows its throughput and latency performance as well as its resilience to hardware and data center failures. This paper presents the design and evolution of ACOS, evaluates its performance in production, and demonstrates its capacity to scale and support Apple’s current storage needs and future growth.
Speaker Bio
Swapnil Pimpale is an Engineering Manager at Apple, working on large-scale distributed storage systems in Apple’s cloud infrastructure. He leads engineering efforts spanning performance/efficiency improvements at massive scale, emerging drive technologies and disk hardware management. Swapnil is a co-author of the USENIX FAST 2026 paper on Apple’s geo-distributed object store and has also published systems research in USENIX venues, including HotStorage. Earlier in his career, he worked on storage and operating systems research connected to Carnegie Mellon University’s Parallel Data Lab (PDL). He is an inventor on multiple issued U.S. patents related to distributed storage and has contributed to the open-source storage community (Lustre).
PolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases
PolarStore is a compressed shared storage system for cloud-native RDBMSs. PolarStore employs a dual-layer compression mechanism that combines in-storage compression in PolarCSD hardware with lightweight compression in software. This design leverages the strengths of both approaches. PolarStore also incorporates database-oriented optimizations to maintain high performance on critical I/O paths. Drawing from large-scale deployment experiences, we also introduce hardware improvements for PolarCSD to ensure host-level stability and propose a compression-aware scheduling scheme to improve cluster-level space efficiency. PolarStore is currently deployed on thousands of storage servers within PolarDB, managing over 100 PB of data. It achieves a compression ratio of 3.55 and reduces storage costs by approximately 60%, and these savings are achieved while maintaining performance comparable to uncompressed clusters
Speaker Bio
Xinjun (Jimmy) Yang is the Senior Director of PolarDB, Database Group, Alibaba Cloud. Leading developing one of the fastest growing databases in Asia Pacific. His recent work won the 2024 and 2025 Best Industry Paper Award on ACM SIGMOD conference, one of the most prestigious and premier conferences for database researchers, engineers, and practitioners. Before joining Alibaba, he worked as InnoDB Architect in Oracle MySQL Team and earlier as a DBAcess group manager in the Sybase Enterprise Server Team.
uCache: A Customizable Unikernel-based IO Cache
Modern storage hardware like NVMe SSDs has outpaced the traditional OS storage stack, leaving developers with a dilemma: stick with simple but slow OS-level caching (mmap), or design complex, custom userspace solutions. In this talk, I will introduce uCache, a novel IO caching architecture that eliminates this trade-off. By leveraging a unikernel-based design, uCache bridges the gap between userspace and in-kernel approaches, enabling efficient and scalable operations that can be customized to the application workload. In particular, unlike kernel-bypass approaches that often require implementing custom storage stacks, uCache's integration into the OS enables it to leverage existing, robust filesystem implementations for file management while optimizing the data path for NVMe performance.
Speaker Bio
Ilya is a 3rd year PhD candidate at the Technical University of Munich. His research interest lies in the design of specialized operating systems that leverage application semantics to improve performance and resource utilization of data-intensive applications. He will be looking for an internship in the industry or for visiting another research group in academia in the coming year.