Milvus 2.1 Timeline
What's inside milvus 2.1?
Functionality:
String Support
MEP 11 -- Support String DataType
Data Time To Live(TTL)
Support automatically purge outdated data
Multi Memory Replica
Speculative execution
Speculative execution is an optimization technique where a computer system performs some task that may not be needed
Architecture Change:
Plugin Message Storage
More message storage options, such as Kafka, pulsar, rockmq....
Add Log node as a new component
each channel will be handled by one log node, all time tick will be sent by log node rather than root coord.
log node can implement log indexes for PK → Segment mapping
Move knowhere out of Milvus as a independent oss project
Performance Boosting:. Goal → 10ms e2e latency under small data size(Minimize the overhead), 10000 + QPS for 100m 768 dim vectors(Maximize Throughput)
Full Performance benchmark
Performance bottle neck removal
In Memory RPC under standalone mode
change grpc into a new implementation
Request combination
Merge multiple small nq request into a large nq request to improve throughput
Ease of use:
Support Milvus Standalone as a single process
Support Milvus as a python embedded library.
Dev && run milvus on Mac and Arm server
Others
Membership
Mentorship
Looking for volunteers on refine technical documentation
Refines logs with new standard.