Autonomous-driving R&D pipelines generate petabytes of multi-modal sensor, simulation, and annotation data each month. Teams therefore need a storage layer that scales with their fleets—without forcing them to rewrite trusted tools.
In this talk we present a cloud-agnostic, open-source elastic file system already running in production at multiple OEMs and robo-taxi startups:

- Single namespace beyond 100 PB / 100 billion files — backed by commodity object storage yet mounted as a local POSIX volume.

- Thousands of Kubernetes nodes (mixed GPU/CPU) share the same dataset for training, validation, and replay with zero duplication.

- Multi-protocol endpoints (POSIX, S3, HDFS, WebDAV) let CV, mapping, and simulation teams keep their existing workflows.

- Hybrid & multi-cloud deployment—public cloud or on-prem with Ceph, enabling fully air-gapped SDV Labs.

We will walk through the high-level design patterns—stateless clients, scale-out metadata, transparent tiering—and share measured outcomes: saved millions in storage cost, scaled single-volume capacity 10×, with painless vendor portability. Attendees will leave with an opinionated reference diagram they can implement immediately.