展开查看详情
1 .The journey of Milvus
金海 Zilliz
2 .Speaker bio
金海
R&D Director
Project Leader, Committer
Career history
Education
© 2020 Milvus. All rights reserved.
3 .Why Milvus
What do we have?
• Faiss, HNSW, Annoy, LSH, NGT, SPTAG, ScaNN ……
What do we need? Vector Search Engine / Vector DBMS
• Supporting different indexes
• Supporting large-scale dataset and different datasets
• Supporting CRUD operations
• Utilizing Computation resources efficiently
• Scalability, High availability
• ……
© 2020 Milvus. All rights reserved.
4 .Challenges – Data Management
Training data
{0.32, 0.21, 0.23, ...... 0.51}
Faiss - IVF {0.52, 0.11, 0.72, ...... 0.60}
Flat / SQ / PQ {. . . . . ..... . .}
{0.91, 0.33, 0.54, ...... 0.13}
Centroids …… Vector data
{0.32, 0.21, 0.23, ...... 0.51}
{0.52, 0.11, 0.72, ...... 0.60}
{. . . . . ..... . .}
{0.91, 0.33, 0.54, ...... 0.13}
……
Clusters
Memor
y
© 2020 Milvus. All rights reserved.
5 .Challenges – Data Management
{0.32, 0.21, 0.23, ...... 0.51}
{0.52, 0.11, 0.72, ...... 0.60}
{. . . . . ..... . .} Segment
{0.91, 0.33, 0.54, ...... 0.13}
Segment Segment
Segment
Segment
Segment Segment with Index
Segment
Collection
<<Milvus在大规模向量检索场景下的数据管理>>
© 2020 Milvus. All rights reserved.
6 .Challenges – Multiple GPU support
Faiss - IndexReplicas Faiss - IndexShards
GPU0 GPU1 GPU0 GPU1
© 2020 Milvus. All rights reserved.
7 .Challenges – Multiple-GPU support
Segment0 Segment1
Segment2 Segment3 11GB/
s
Segment4 Segment5 Segment3
Segment0
Segment4 Segment1
Segment2
Segment5
CPU Memory GPU0 memory GPU1 Memory
<<Milvus 向量搜索引擎：查询任务调度原理>>
© 2020 Milvus. All rights reserved.
8 .Challenges – CPU-GPU co-
processing
Faiss – IVF Query
Index {0.32, 0.21, 0.23, ...... 0.51}
Step1 Centroids ……
Query
{0.32, 0.21, 0.23, ...... 0.51}
Step2 ……
Clusters
© 2020 Milvus. All rights reserved.
9 .Challenges – CPU-GPU co-
processing
Faiss – IVF Index IVF-SQ8-Hybrid
nlist
2x-5x speed up!
nlist * dimension * 4
Centroids …… GPU
bytes
……
CPU
Clusters
© 2020 Milvus. All rights reserved.
10 .Challenges – Entity Deletion
Faiss – IVF
Index
Centroids ……
……
Clusters
© 2020 Milvus. All rights reserved.
11 .Challenges – Entity Deletion
Milvus – IVF Index
Centroids ……
bitmap bitmap bitmap bitmap
……
Clusters
<<Milvus数据删除的实现原理>>
© 2020 Milvus. All rights reserved.
12 .Milvus Releases
v0.1 v0.2 v0.3 v0.4 v0.5 v0.6
C++ Implemented Supports IP metric Milvus CPU-only
Implementation Adds Multiple-GPU Supports IVF-SQ8H
Supports data loading Supports IVF SQ Supports IVF-PQ
based on Faiss support index
and computation index Index
Python APIs Adds Java SDK
pipeline Mishards open
sourced
v0.7 v0.8 v0.9 v0.10 v0.11
Supports entity Supports Annoy Index Supports selecting CPU Query request merge Scalar Field Filtering
deletion instruction set during Supports large topk Meta Snapshot
runtime search
© 2020 Milvus. All rights reserved.
13 .0.11.0 – Scalar Field Filtering
Milvus
DSL AST AST
Parser Validator Optimizer
Listener Plan
Request
Query
Client DSL Context Executor
Connection
Connection
Connection Resul
Context
Context
Response Context t
Resul Data
t
Cache
Data
Data on disk
© 2020 Milvus. All rights reserved.
14 .0.11.0 – Meta Snapshot
New Merg
Q1 Req Q2 Req Q3 Req Q1 Rep Q2 Rep Q3 Rep
Segment e
Timeline
Collection A Snapshots
Segment: S1
Versio
Q1 Segments
n Latest snapshot
Segment: S2 Q1 v1 S1, S2
Latest snapshot
Q2
Q2 v2 S1, S2, S3
Segment: S3 Latest snapshot
Q3 v3 S3, S4
Q3 Segment: S4
© 2020 Milvus. All rights reserved.
15 .Next Step
• Milvus Technical Steering Committee
• How to contribute
• Milvus github Discussions (https://github.com/milvus-io/milvus/discussions)
© 2020 Milvus. All rights reserved.
16 .Fin
Speaker Contact
WeChat: 13917405881
Mail: hai.jin@zilliz.com
热门城市
