Major functionalities
Change data capture
User can subscribe both DDL and DML data from the message storage
Note: the v2.2 is delayed to 2022.10.30
Stability improvement
Stabilitiy is the first priority of milvus for now.
The community decide to improve stability to a new step in the next major release, including:
- Refactor Rootcoord to support idempotent retry of all DDL actions
- Refactor Querycoord to solve the issues of loading collection error and failure recovery
- Refactor IndexCoord and refine index build process
- Support coordinator active/standby switch
- Increase the UT coverage for all coordinator and Knowhere execution engine to 90%
- Figure out some major memory issues.
- Add fully e2e test for Java and Golang SDKs.
- Support Quota to avoid overload the cluster
Major functionalities
Search by ID
Given a entity ID, search for the closest TOP N of this entity
Segment LRU load
If data/index size is too large to fit in memory, support to load segment in LRU mode from disk.
Support alter collection attribute/schema
Alter collection to update attributes such as replica number, compression or add one field to collection
Vector Range Search (search by distance)
...
Support Faiss GPU index building and search.
Disk INDEX
Support DiskANN as on disk index
RBAC Support
Role based access control, See design of MEP29
Data backup
Support data backup and restore, help you to keep your data safe
Enhancement
Refactor collection Alias
Support atomic switch of collection alias.
Refactor meta storage and meta storage format
Support mysql meta storage for some of the large metas, change meta storage format.
Performance optimization
Performance boost on scalar filtering and vector comparison
Others
Observability
Refine Log and Error message
Support more vector index such as diskANN(Optional)
Auto Scalar(Optional)Monitoring
Support open tracing in Milvus