[gaius-qi]

Description

Makes the image APIs of client-request (Go and Rust) image-aware. PreheatImage, StatImage and DeleteImage now all talk to the seed peers directly; the scheduler is only used to list the seed peers.

sequenceDiagram
    participant SDK
    participant Registry
    participant SeedPeers as Seed peers (dfdaemon)

    Note over SDK,SeedPeers: PreheatImage
    SDK->>Registry: resolve manifests (all platforms, or one)
    SDK->>SeedPeers: DownloadTask(url, image{reference, platform})

    Note over SDK,SeedPeers: StatImage
    SDK->>Registry: resolve manifests (all platforms, or one)
    SDK->>SeedPeers: StatLocalTask(task id) for every layer

    Note over SDK,SeedPeers: DeleteImage
    SDK->>SeedPeers: DeleteImage(manifest url, platform)
    SeedPeers-->>SDK: deleted / retained task ids

Loading

API	Before	After
PreheatImage	Preheats one platform, the current platform by default.	Preheats all the platforms unless a platform is set. Every download carries the image reference and the platform, so dfdaemon records which image a task belongs to.
StatImage	Calls the scheduler StatImage gRPC, which needs the scheduler job service (redis).	Resolves the image itself and asks every healthy seed peer for each layer with dfdaemon StatLocalTask.
DeleteImage	Resolves the image from the registry and deletes every manifest and blob task from its replicas.	Calls dfdaemon DeleteImage on every healthy seed peer, without resolving the image. A task still referenced by another image or platform is retained. Returns the deleted and retained task ids of each seed peer.

API changes

 // Go (the Rust crate mirrors it with the same names in snake_case)
-DeleteImage(ctx, req) error
+DeleteImage(ctx, req) (*DeleteImageResponse, error)   // Peers []*PeerDeletedImage{IP, Hostname, DeletedTaskIDs, RetainedTaskIDs}

 // DeleteImageRequest options
+WithDeleteImageRequestConcurrentPeerCount             // default 500
 WithDeleteImageRequestPlatform
 WithDeleteImageRequestTimeout
-WithDeleteImageRequest{Auth,PieceLength,Tag,Application,FilteredQueryParams,
-  ContentForCalculatingTaskID,EnableTaskIDBasedBlobDigest,Replicas,Scope,ConcurrentTaskCount}

 // StatImageRequest options
+WithStatImageRequestConcurrentPeerCount               // default 500

-PeerImage
+PeerStatedImage

• The platform is optional (*string / Option<String>), unset means all the platforms, and it is validated: containerd/platforms in Go, the os/arch form in Rust.
• StatImage and DeleteImage fan out to the seed peers with a concurrency limit. A seed peer that fails is left out of the response instead of failing the request.
• An image index that lists the same manifest twice for a platform is preheated once.

Notes

• The seed peers must run dfdaemon v1.5.12 or later. An image preheated by an earlier SDK has no image record, so DeleteImage deletes nothing for it.
• StatImage now needs the registry to be reachable from the SDK, as PreheatImage does.
• With a platform set, Go follows the strict matcher of containerd (linux/arm selects arm/v7 only) and normalizes aliases such as linux/aarch64; Rust selects every index entry with that os/arch and sends the platform as given.
• Bumps d7y.io/api/v2 to v2.3.10 and d7y.io/dragonfly/v2 to v2.5.4-rc.0, dragonfly-api to 2.3.10, and the crate version to 1.13.0. The e2e cluster uses scheduler v2.5.4-rc.0 and client v1.5.12.

Tests

• Go: table-driven tests for PreheatImage, StatImage and DeleteImage against an in-test TLS registry and mock seed peers (targets and Download.image per platform, cached layers per seed peer, a failed seed peer left out, retries, invalid arguments).
• Rust: table-driven twins for delete_image and for the invalid arguments of the three APIs. The success paths of preheat_image and stat_image need a registry and are covered by the e2e tests.
• e2e: the drivers gain preheat-image --platform, stat-image and delete-image --platform. The suite grows from 31 to 53 specs, run for both SDKs: preheat all platforms, one platform and by digest; stat by platform and after delete; delete by platform, platforms preheated one by one, a platform or an image that is not preheated, a tag and a digest reference sharing a platform, two tags sharing one blob, and an image preheated by one SDK and deleted by the other.

Unit tests pass locally (go test -race, cargo test --all-features, clippy with and without the preheat feature). The e2e specs run in CI.

Related Issue

• feat: preheat and delete all the platforms of the image by default api#698
• feat: delete the image tasks not referenced by other images client#2159
• Enhance multi-platform image preheat and deduplication process dragonfly#5112

Motivation and Context

Deleting an image by resolving its manifest removes blobs that other images still share, and fails when the registry is unreachable or the tag has moved. dfdaemon v1.5.12 records the image and the platform every task is downloaded for, so the SDK passes them at preheat and lets dfdaemon decide what to delete.

StatImage went through the scheduler job service, which only exists when redis is deployed. Asking the seed peers directly makes it work like PreheatImage and DeleteImage in a scheduler-only deployment.

Screenshots (if appropriate)

N/A

🤖 Generated with Claude Code