Is there a benefit to doing CoW with Pandas vs. offloading it to the storage? Practically all modern storage systems support CoW snaps. The pattern I’m used to (Infra, not big data) is to leverage storage APIs to offload storage operations from client systems.