Milvus Build Up the Unstructured Data Service - Jun Gu 09.2020 - ITU
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Milvus
Build Up the Unstructured Data Service
Jun Gu
09.2020
© 2020 Zilliz. All rights reserved.
Speaker bio
Jun Gu
Database engineer, SME
Voting member in Technical Advisory Council (TAC)
Partner, Chief Evangelist
Career history
Education
© 2020 Zilliz. All rights reserved.
Zilliz: Who we are
• Open source software company based
in Shanghai
• Mission: Reinvent data science
• Main contributor of Milvus project
© 2020 Zilliz. All rights reserved.
Unlock the treasure of unstructured data
AI algorithms transform image, video, voice, natural language into vectors,
and enables understanding and utilization of unstructured data at scale.
Unstructured data Deep learning models Vectors Knowledge, insight, $
© 2020 Zilliz. All rights reserved.
The flow-based AI applications
Video
Extract Voice Extract The most popular way
frames model tags
• Flexible
• Easy to compose, web-based UI
• Sample piplelines
Image
The challenge
Visual model • Data fragmentation
VGG, eg.
Vectors Vectors
Attributes
Visual Voice
© 2020 Zilliz. All rights reserved.
The sample pipelines for video processing
The unstructured data service (UDS) for AI
Unstructured Data
image, video, voice, natural language
Search Insert
Model Inference Runtime
store Inference Layer
TensorRT, ONNX RT, TFRT
Search Insert
Milvus
Vectors Attributes Object
Data Service Layer
High dense + Sparse (will be in 0.11) Storage
Multimodal Scoring
(will be in 0.14) (will be in 0.16)
output
Result Set
image, video, voice, natural language
© 2020 Zilliz. All rights reserved.
Why Milvus: Vectors are different
Numbers Vectors
Arithmetic operation Similarity (eg. Euclidean distance)
d ( A, B )
n
(a b )
i i
2
Operation i 1
Number comparison Similarity comparison
a b
TopK ( A) arg min(d ( A, B ))
B
1–10
1–5 6–10 Organization
1 2 3 4 5 6 7 8 9 10
© 2020 Zilliz. All rights reserved.
Milvus: The big picture
Query Scheduler Processing Engine Buffer Pool
ANNS Collaborative Query
Mi-FAISS, Mi-Annoy tag/structured data Index
Result Files
SDK / Web API
top-K result
Reducer Multi-modal
Scoring
app specific
Segment Segment
query obj Metadata
Selection
insert obj
X86: supports SSE4.2, AVX2, AVX512
GPU: Pascal microarchitecture or later, CUDA 10.0
or later x86 ARM GPU New Index
Arm: requires aarch64 Index
Kunpeng: tested on Kunpen 920 with Centos 7.x Files
File
Loongson: tested on Loongson with docker
container Kunpeng Loongson RISC-V
RSIC-V: in early development
Various Processors Storage Tier
© 2020 Zilliz. All rights reserved.
Milvus: Distributed deployment
© 2020 Zilliz. All rights reserved.
Milvus: The ANN benchmark
Milvus: 0.8.0
OS: Ubuntu 18.04
ECS: AWS c5.4xlarge (16c, 32GB), Intel XeonPlatinum 8275CL
Data set: sift-128-euclidean (1 million vectors)
More info: https://milvus.io/docs/benchmarks_aws
© 2020 Zilliz. All rights reserved.
Special thanks to ANN-Benchmarks (developed by Martin Aumueller, Erik Bernhardsson and Alec Faitfull)Milvus: The journey
2018.10 2019.04 2019.06
The most active AI projects in
1st
The Milvus Linux foundation
seed
idea 0.1
user
Open Joined
Source LF AI
2019.10 2020.03
© 2020 Zilliz. All rights reserved.Progress
Unstoppable momentum since its debut.
5.9K 3.9K 104
Commits GitHub stars Contributors
16 200+ 19
Release Users Patents filed
© 2020 Zilliz. All rights reserved.Comprehensive Leading-Edge Dynamic Data
Similarity Metrics Performance Management
Milvus
Features & benefits Near Real Time
Rich Data Type
& Advanced Cost Efficient
Search
Search
The world’s most advanced, our target
Highly Scalable
and Robust Cloud Native Ease of Use
© 2020 Zilliz. All rights reserved.Use case: Inteligent writing assistant
Corpus Data
natural language
Writing Intention
Data Cleansing
Feature engineering
Encoder
TextCNN
Extract paragraph, summary
Result
An auto-generated
Encoder essay
InferSent
Object
Milvus
Storage
© 2020 Zilliz. All rights reserved.Use case: News recommendation on mobile
Daily batch Feeding
News title News title
Encoder
SimBert
Object
Milvus
Storage
Reading Recommended
Preference News
© 2020 Zilliz. All rights reserved.Use case: Image search for company trademark
Images
Company Trademark • 55 million images
• Search elapsed time:
20 ms on cloud GPU server
Encoder
VGG (fine tuned)
Search Object
Milvus
Storage
Trademark Image Company Info
© 2020 Zilliz. All rights reserved.Use case: Pharmaceutical molecule analysis
Molecular Formula
• 800 million molecules
CC(=O)Nc1ccc(S(=O)(=O)NCC(=O)N2CCS(=O)CC2)cc1
• Search elapsed time:
Encoder
500 ms on single server RDKit
Molecular fingerprint: 1024 bits
00001100...10000000
Milvus
Tanimoto similarity
Molecular
Substructure Candidate List
Similarity
© 2020 Zilliz. All rights reserved.
SuperstructureUseful Links
Performnance benchmark:
https://milvus.io/docs/benchmarks_aws
https://milvus.io
https://github.com/milvus-io/milvus Live demo:
https://milvus.io/scenarios
https://milvusio.slack.com
https://twitter.com/milvusio • Content-based image retrieval system (以图搜图)
https://medium.com/unstructured-data-service
• Q&A chatbot powered by NLP (智能客服机器人)
• Molecular analysis (化合物分析)
https://zhuanlan.zhihu.com/ai-search
© 2020 Zilliz. All rights reserved.Thanks!
© 2020 Zilliz. All rights reserved.You can also read
