mirror of
https://github.com/milvus-io/milvus.git
synced 2026-10-12 02:19:10 +00:00
wei liu
11737297b0
issue: #52575 ## What changed - Removed DataNode-version filtering from copy-segment scheduling. - Kept local restore on `CopySegment` and routed external restore through `ExternalCopySegment`. - Retained payload-based external detection for older DataCoord clients. - Made DataNode return `ServiceUnimplemented` for unknown task types. - Normalized the exact legacy code-5 rejection for `ExternalCopySegment` to the same unsupported-task error. - Failed external restore immediately when the worker does not support the task type, while keeping other worker errors retryable. - Retained `QuerySlotResponse.version` for wire compatibility without using it for worker capability selection. ## Why Cloud worker endpoints can be pooled, so the version returned by a slot endpoint does not reliably identify the DataNode that executes the task. Version-based filtering can therefore exclude compatible workers indefinitely. The task type now acts as the capability boundary. New workers return a typed unsupported error for unknown task types. DataCoord also recognizes the exact legacy rejection returned by older workers, so an incompatible external restore fails immediately instead of being requeued until the job timeout. Transient worker failures remain retryable. Local restore keeps its existing task type and behavior. This change intentionally does not add failed-worker exclusion or cross-worker rerouting to the scheduler. ## Validation - Rebasing onto `milvus/master` preserved the complete feature patch ID and changed-file set without conflicts. - Targeted race tests passed after the rebase for `pkg/taskcommon`, `pkg/util/merr`, `internal/datacoord/session`, `internal/datacoord`, and `internal/datanode` with `dynamic,test` tags and disabled inlining. - The tests cover typed unknown-task rejection, legacy RPC status round-trip, DataCoord fail-fast behavior, and retry preservation for unrelated errors. - `git diff --check` passed. The post-rebase local C++ build is blocked by an unrelated `milvus/master` macOS/clang 19 error in `ConcurrentVector.cpp`, where `std::min` receives `int64_t` and `ssize_t`. This PR does not modify `internal/core`. --------- Signed-off-by: Wei Liu <[email protected]>