podman save --format oci-archive: loses image ID

[edsantiago]

This may not be a bug, but I think it is, because podman save without --format oci-archive preserves image ID.

$ cat >oci-check.sh <<EOF
#!/bin/bash

image=${1?Usage: $0 FQIN}

OCI_FORMAT="--format oci-archive"

./bin/podman rm  -a >/dev/null
./bin/podman rmi -a >/dev/null
./bin/podman pull -q "$image" >/dev/null
./bin/podman images --noheading
rm -f /tmp/foo.tar
./bin/podman save $OCI_FORMAT -o /tmp/foo.tar "$image"
./bin/podman rmi -a >/dev/null
./bin/podman load -q -i /tmp/foo.tar   >/dev/null
./bin/podman images --noheading
EOF
$ chmod 755 ./oci-check.sh

!! basically just "podman pull", "podman save", and "podman load". Note differing image IDs.
$ for i in docker.io/library/busybox docker.io/library/alpine:latest quay.io/libpod/alpine_labels:latest;do ./oci-check.sh $i;echo;done
docker.io/library/busybox  latest  018c9d7b792b  3 weeks ago  1.45 MB
docker.io/library/busybox  latest  c71b90eb5b17  3 weeks ago  1.44 MB

docker.io/library/alpine  latest  a24bb4013296  2 months ago  5.85 MB
docker.io/library/alpine  latest  0f5f445df8cc  2 months ago  5.85 MB

quay.io/libpod/alpine_labels  latest  4fab981df737  21 months ago  4.69 MB
quay.io/libpod/alpine_labels  latest  dd8b0a90b986  21 months ago  4.68 MB

Commenting out OCI_FORMAT= makes all the IIDs agree.

master @ 7e2a1b3

[vrothberg]

@mtrmac, do you think it's a bug or feature request/enhancement?

[mtrmac]

In general, neither. The change ultimately happens in https://github.com/containers/image/blob/1c313b2d23e039a81a64bbced1b2b427d60429e1/image/docker_schema2.go#L78 , and as the comment says, the conversion from v2s2 to OCI is lossy (IIRC OCI does not have health checks).

So, workflows that convert from v2s2 to OCI should be prepared for the config digest == image ID changing. (If the missing fields are the only difference between the formats, we could check that nothing of substance has changed when doing the conversion, and in that case use the original raw bytes to preserve the ID. I don’t think that’s desirable: it’s extra code to preserve an illusion of a property that does not hold in general.)

[rhatdan]

@edsantiago Does this answer your question?

[edsantiago]

Yes, it does. Thanks @mtrmac.

[vrothberg]

I entirely oversaw the conversion, thanks for clarifying @mtrmac!