ATC 2026

Abstract

For fifty years, storage systems have kept one promise: once written, data is never lost. Journaling, logging, and replication were built to keep it, and fsync() became the interface through which applications ask for it. File systems extend this promise to nearly everything they store, from bank records to temporary files, because they cannot tell what the data is worth. Keeping the promise has never been free, and much of our research, including my own, has tried to make it cheaper.

As flash replaced disks, the cost of enforcing write ordering through durability came to dominate I/O latency, which led us to separate the two guarantees. As core counts grew, the journal commit became a serial point that limited the whole machine, which led us to rethink how transactions are committed and what unit they should cover. As log-structured designs spread, garbage collection became the price of never overwriting, and we had to ask who should pay it and where the time really goes. In each case, the largest gains came from questioning a long-held assumption rather than from tuning.

Yet we have rarely asked which data needs the durability guarantee at all. In the era of agentic AI, the assumptions that have governed storage for fifty years no longer hold. The primary client of storage is shifting from the CPU to the GPU, and this shift raises a fundamental question: is durability still what storage is for? In this talk, I look back at what the I/O stack has taught us, revisit the durability guarantee that storage systems have strived to uphold for decades, and discuss how the role of storage changes in the era of agentic AI.