Chris Evich

Bump: Ran into this fairly hard when working on https://github.com/containers/podman/discussions/19796 Scenario: You need to sync a huge number of images across multiple registry namespaces. It breaks somewhere in the middle...

Related: https://github.com/containers/skopeo/issues/1240#issuecomment-878472089 Podman is falling back to `/tmp/podman-run-1000/containers/auth.json`, buildah seems to behave similarly but I can't confirm. The debug output from both reports using `containers-auth.json` (which is wrong).

So I poked around in the podman vs skopeo code. The best I can conclude is the behavior comes from differences in setup of the `sys` (called `sysCtx` in podman)...

Are you sure this is coming from `/run/containers`? The behaviors I observed *only* happen in skopeo and not podman or buildah. In my mind that means the fix needs to...

Right, okay I think I follow. Also moving the `c/image` default would empower all of buildah, skopeo, and podman to "follow the same rules" instead of each doing something different...

> Podman and Buildah are using contaers/common/pkg/auth By my reading, skopeo is doing it that way too, but the setup of options passed in to `auth.Login()` is different. I'm guessing...

IIRC (it's been a loooooong time, memory is fuzzy), podman has very good reasons for guaranteeing XDG_RUNTIME_DIR is set, and buildah also. I my mind, this might be reason for...

[Oh hey, look at that, there's already a c/common issue open on this.](https://github.com/containers/image/issues/1125)

Checking: We're not sure, but Miloslav suggested I may be seeing this issue when building multiarch images (using buildah) in Cirrus-cron jobs. For whatever reason, it tends to happen once/twice...

> The might be applicable if there are two processes, one is pulling the image, and another is referring to it (using a known top-level digest?) and finds the not-yet-complete...