mirror of
https://gitee.com/milvus-io/milvus.git
synced 2026-01-07 19:31:51 +08:00
f9827392bb
issue: #45881 Add persistent task management for external collections with automatic detection of external_source and external_spec changes. When source changes, the system aborts running tasks and creates new ones, ensuring only one active task per collection. Tasks validate their source on completion to prevent superseded tasks from committing results. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> - Core invariant: at most one active UpdateExternalCollection task exists per collection — tasks are serialized by collectionID (collection-level locking) and any change to external_source or external_spec aborts superseded tasks and causes a new task creation (externalCollectionManager + external_collection_task_meta collection-based locks enforce this). - What was simplified/removed: per-task fine-grained locking and concurrent multi-task acceptance per collection were replaced by collection-level synchronization (external_collection_task_meta.go) and a single persistent task lifecycle in DataCoord/Index task code; redundant double-concurrent update paths were removed by checking existing task presence in AddTask/LoadOrStore and aborting/overwriting via Drop/Cancel flows. - Why this does NOT cause data loss or regress behavior: task state transitions and commit are validated against the current external source/spec before applying changes — UpdateStateWithMeta and SetJobInfo verify task metadata and persist via catalog only under matching collection-state; DataNode externalCollectionManager persists task results to in-memory manager and exposes Query/Drop flows (services.go) without modifying existing segment data unless a task successfully finishes and SetJobInfo atomically updates segments via meta/catalog calls, preventing superseded tasks from committing stale results. - New capability added: end-to-end external collection update workflow — DataCoord Index task + Cluster RPC helpers + DataNode external task runner and ExternalCollectionManager enable creating, querying, cancelling, and applying external collection updates (fragment-to-segment balancing, kept/updated segment handling, allocator integration); accompanying unit tests cover success, failure, cancellation, allocator errors, and balancing logic. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Signed-off-by: sunby <sunbingyi1992@gmail.com>
Data Coordinator
Data cooridnator(datacoord for short) is the component to organize DataNodes and segments allocations.
Dependency
- KV store: a kv store has all the meta info datacoord needs to operate. (etcd)
- Message stream: a message stream to communicate statistics information with data nodes. (Pulsar)
- Root Coordinator: timestamp, id and meta source.
- Data Node(s): could be an instance or a cluster, actual worker group handles data modification operations.