A label selector now freezes the Postgres data and WAL volumes at the same instant

A VolumeGroupSnapshot carries a label selector. A selector like `app=postgres` picks up the data PVC and the logs PVC together ^[14]. The New Stack calls that the important design choice, because the group boundary is then expressed in Kubernetes terms that survive adding or resizing volumes over time ^[14]. Enumeration cannot do that. A tool that walks the PVC list at backup time is inferring an application boundary the per-PVC model never recorded ^[3]. Operators used to declare that boundary by hand, telling a storage array which LUNs belonged to the same application ^[4].

Each per-PVC snapshot is crash-consistent on its own, equivalent to pulling the power cord on that one volume, and none of them is consistent with the next ^[6]. The New Stack puts the interval between two sequential snapshots at a few hundred milliseconds of continued writes ^[7]. Take that at 300 ms and a three-volume application leaves the first snapshot about 600 ms behind the last, because the window is the sum of the gaps ^[18]. It widens with write rate and with volume count ^[8].

Whether that makes last night's backup worthless turns on something the version bump leaves alone. Quiescing closes the window: freeze I/O, flush buffers, snapshot, unfreeze ^[10]. Teams doing that already have coherent sets. Teams that ruled it out did so because freezing a busy database for the duration of a multi-volume snapshot is the disruption backups exist to avoid ^[10]. They have been producing what The New Stack describes as "a set of snapshots that each looks healthy and collectively describes a state that never existed" ^[9]. On the 2 a.m. Postgres restore, The New Stack wrote: "The backup wasn't corrupted in transit. It was inconsistent the moment it was taken." ^[17]

Adoption is three objects and one dependency. An administrator defines a VolumeGroupSnapshotClass describing how group snapshots are created for a given CSI driver, the user creates the VolumeGroupSnapshot with its selector, and a VolumeGroupSnapshotContent tracks the provisioned result ^[11]. The atomic part belongs to the driver: "Under the hood, the CSI driver takes one atomic, point-in-time snapshot across every selected volume," according to The New Stack, with no application quiescence required, provided the underlying storage supports it ^[12][13].

So general availability in v1.36 is a promotion of the API ^[1]. In my view the presumption worth carrying is a narrow one: a multi-volume set taken without quiescence has never been shown to restore, and showing it costs one restore drill ^[16]. That drill is also the only point at which the gap becomes visible.