갤럭시 규모의 인공지능 서비스를 위한 AWS 데이터베이스 아키텍처 - 김상필 솔루션 아키텍트 매니저, AWS / 김정환 데브옵스 엔지니어, 삼성전자 / 이혁 데이터베이스 운영 엔지니어, 삼성전자 :: AWS Summit Seoul 2019
3 likes•1,962 views
The document presents an overview of AWS database architecture for large-scale AI applications, particularly focusing on Samsung's Bixby 2.0 platform. It highlights the characteristics and data requirements of modern applications, the purpose-built database options available in AWS, and specific use cases involving DynamoDB and Aurora for optimal performance. Additionally, it discusses the challenges faced in previous versions of Bixby and how the new architecture addresses these issues while enabling scalability and efficient data management.
갤럭시 규모의 인공지능 서비스를 위한 AWS 데이터베이스 아키텍처 - 김상필 솔루션 아키텍트 매니저, AWS / 김정환 데브옵스 엔지니어, 삼성전자 / 이혁 데이터베이스 운영 엔지니어, 삼성전자 :: AWS Summit Seoul 2019
- 1. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. 갤럭시 규모의 인공지능 서비스를 위한 AWS 데이터베이스 아키텍처 김상필 솔루션즈 아키텍트 매니저 AWS Korea 이혁 Database Engineer Samsung Electronics 김정환 Global Service Reliability Engineer Samsung Electronics
- 2. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. 발표자료 바로 공개 발표자료는 발표 종료 후 해당 사이트에서 바로 보실 수 있습 니다 © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 3. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. What are modern applications? AWS Database service options Case study: Samsung Bixby 2.0 Case study: Bixby Database architecture Agenda
- 4. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Characteristics of modern applications Internet-scale and transactional Users: 1M+ Data volume: TB–PB–EB Locality: Global Performance: Milliseconds–microseconds Request Rate: Millions Access: Mobile, IoT, devices Scale: Up-out-in Economics: Pay-as-you-go Developer access: Instant API accessSocial mediaRide hailing Media streaming Dating
- 5. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Modern applications require internet-scale performance
- 6. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Data categories and common use cases Relational Key value Document In-memory Graph Referential integrity, ACID transactions, schema- on-write Low-latency, key lookups with high throughput and fast ingestion of data Indexing and storing documents with support for query on any attribute Microseconds latency, key- based queries, and specialized data structures Creating and navigating data relations easily and quickly Lift and shift, EMR, CRM, finance Real-time bidding, shopping cart, social Content management, personalization, mobile Leaderboards, real-time analytics, caching Fraud detection, social networking, recommendation engine Search Indexing and searching semistructured logs and data Product catalog, help, and FAQs, full text Time series Ledger Collect, store, and process data sequenced by time IoT applications, event tracking Complete, immutable, and verifiable history of all changes to application data Systems of record, supply chain, healthcare, registrations, financial
- 7. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T AWS: Purpose-built databases Relational Key value Document In-memory Graph Search Amazon DynamoDB Amazon Neptune Amazon RDS Aurora CommercialCommunity Amazon ElastiCache Amazon Elasticsearch Service Amazon DocumentDB Time series Ledger Amazon Timestream Amazon Quantum Ledger DatabaseMemcachedRedis
- 8. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Amazon DynamoDB Fast and flexible key value database service for any scale Comprehensive security Encrypts all data by default and fully integrates with AWS Identity and Access Management for robust security Performance at scale Consistent, single-digit millisecond response times at any scale; build applications with virtually unlimited throughput Global database for global users and apps Build global applications with fast access to local data by easily replicating tables across multiple AWS Regions Serverless No hardware provisioning, software patching, or upgrades; scales up or down automatically; continuously backs up your data
- 9. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Amazon Aurora MySQL and PostgreSQL compatible relational database built for the cloud Performance and availability of commercial-grade databases at 1/10th the cost Availability and durability Fault-tolerant, self-healing storage; six copies of data across three AZs; continuous backup to S3 Fully managed Managed by RDS: no hardware provisioning, software patching, setup, configuration, or backups Highly secure Network isolation, encryption at rest/transit Performance and scalability 5x throughput of standard MySQL and 3x of standard PostgreSQL; scale-out up to 15 read replicas
- 10. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T When to Use Which Services Situation Solution Existing application Use your existing engine on RDS • MySQL Amazon Aurora, RDS for MySQL • PostgreSQL Amazon Aurora, RDS for PostgreSQL • MariaDB Amazon Aurora, RDS for MariaDB • Oracle Use SCT to determine complexity Aurora, RDS for Oracle • SQL Server Use SCT to determine complexity Aurora, RDS for SQL Server New application • If you can avoid relational features DynamoDB • If you need relational features Amazon Aurora In-memory store/cache • Amazon ElastiCache Time series data • Amazon Timestream Track every application change, crypto verifiable. Have a central trust authority • Amazon Quantum Ledger Database (QLDB) Don’t have a trusted central authority • Amazon Managed Blockchain Data Warehouse & BI • Amazon Redshift, Amazon Redshift Spectrum, and Amazon QuickSight Adhoc analysis of data in S3 • Amazon Athena and Amazon QuickSight Apache Spark, Hadoop, HBase (needle in a haystack type queries) • Amazon EMR Log analytics, operational monitoring, & search • Amazon Elasticsearch Service and Amazon Kinesis
- 11. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Architect services ground-up for the cloud and for the explosion of data Offer a portfolio of purpose-built services, optimized for your workloads Help you innovate faster through managed services Our approach Provide services that help you migrate existing apps and databases to the cloud
- 12. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. 김정환(Jay Kim) Global Service Reliability Engineer Samsung Electronics [email protected] rpkim.net linkedin.com/in/rpkimjay
- 13. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Introduction video
- 14. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Samsung’s A.I. service 2018 Bixby 2.0 2017 Bixby 1.0 NOW 2017 2018
- 15. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T 2019 Bixby 2.0 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant DEMOCRATIZED AI UNPRECEDENTED REACH POWERFUL PERSONALIZATION An open artificial intelligence platform https://bixbydevelopers.com https://bixby.developer.samsung.com Samsung device ecosystem Learn user preferences Bixby 2.0
- 16. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Multi-Locales Service ko-KR en-GB en-US zh-CN de-DE It-IT fr-FR es-ES World-wide Service 60 Million Users Multi-Devices Service Mobile, Television and so on 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0
- 17. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T True benefits of using for 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0
- 18. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Bixby’s Global-scale footprints with AWS ap-northeast-2 (Seoul) cn-north-1 (Beijing) eu-west-1 (Ireland) us-east-2 (Ohio) 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby Production regions Bixby 2.0
- 19. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Bixby’s Global-scale footprints with AWS ap-northeast-2 (Seoul) cn-north-1 (Beijing) eu-west-1 (Ireland) us-east-2 (Ohio) 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com us-west-2 (Oregon) us-east-1 (N. Virginia) cn-north-2 (Ningxia) Bixby 2.0 Bixby Dev/Staging/Production regions
- 20. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Excellent compatibility with third-party tools Packer Spinnaker Ansible Application Provisioning Continuous Delivery Monitoring System Datadog Terraform Infrastructure creation DEV STG Prod 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby using third-party tools Bixby 2.0
- 21. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Excellent compatibility with third-party tools Terraform Infrastructure creation 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby using various continuous delivery pipelines Bixby 2.0 Userdata.sh $ curl http://169.254.169.254/latest/meta-data/ $ role=ABC # from instance tag $ ansible-pull –e “role=ABC“ –U [email protected]:ansible.git playbook.yml Instance Tags ansible-role ABC package-version ABC-273a81-v1.5.zip AnsibleEC2 Ansible # Add cronjob for ansible-pull # Getting package-version # Install & Service Execution Github Infrastructure provisioning : creation machine, infrastructure Application bootstrapping: install & configuration the application
- 22. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Excellent compatibility with third-party tools Bixby using various AWS services for automation Bixby 2.0Security Group Inbound change CloudTrail Detect! CloudWatch Event RevokeSecurityGroupIngress AuthorizeSecurityGroupIngress Lambda Slack EC2 Block Bitcoin mining B
- 23. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T AWS services powering Bixby Compute Container Registry Lambda Elastic Load Balancer Auto Scaling Elastic Container Service EC2 Database Aurora DynamoDB ElasticCache RDS Developer Tools CodeDeploy CLI Security, Identity & Compliance IAM KMS Certificate Manager Storage EBS EFS S3 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0 Management & Governance CloudWatch CloudTrail Trusted Advisor Networking & Content Delivery CloudFront Route53 VPC API Gateway
- 24. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T Why run Artificial Intelligence on AWS 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0 GPU Instances EFS File system Database services For ASR For sharing model data For personal language model
- 25. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T High Level Diagram 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0
- 26. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T PLM Bixby Platform Overview PLM Personal Language Model 2019 AI Interface AI Developer Platform Bixby developer center voice-activated personal assistant https://bixbydevelopers.com https://bixby.developer.samsung.com Bixby 2.0 ASR NLU TTS Storage Services Users Database Services Automatic Speech Recognition ASR NLU Natural Language Understanding TTS Text to Speech
- 27. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Bixby Database architecture 이혁 Database Engineer Samsung Electronics
- 28. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Launching Bixby Bixby 1.0 in 2017 Commercial shared file system Bixby 2.0 in 2018 Database
- 29. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Challenges in Bixby 1.0 • Each personal data started getting bigger • 100MB+ for each device : file loading time • Many small files • Hard to work together • Each team had their own files • Hard to extract meaningful value • Operating File System Cluster on AWS • Disk / Node failure • inode limitations • Managing commercial license
- 30. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. What we needed? Database
- 31. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. What we needed? Blazingly fastPB scalability No license China region
- 32. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. How we selected PLM DB? RDBMS NoSQL Open source Aurora Cloud native Open source Other cloud DynamoDB Cloud native PB scalability
- 33. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. PLM DB PoC scenario • Requirements • No global replication • Database in 1 region + cache in all : network latency • Same architecture for all regions • Each region has 1,920+ concurrent users for load • 300 items per device : 192KB • Generated test data : 10TB • Goal • Get 300 items from DB within 100ms • Put it into Cache once and get 10x within 10ms Seoul OhioIreland DB API API API
- 34. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. PLM DB PoC scenario • Test schema in DynamoDB Device id (PK*) Seq_num (SK**) Last name First name Nick name Full name … x2019 1 Cho EunAh Frozen Eunah Cho … x2019 2 Cho BeomSeung Iron man BeomSeung Cho … x2019 3 Lee Hyuk Groot Hyuk Lee … … * PK : Partitioning Key ** SK : Sort Key
- 35. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. PLM DB PoC result-1 How we tested? 1. API servers in 3 regions called. 2. DynamoDB in Seoul for 1 hour. 3. DynamoDB returned 300 items(192KB). 4. Logged the query + round trip time. 5. Measured average number. Why we did this test? 1. Whether the centralized database was possible or not. 52ms (Seoul) 1,512ms (Ireland) 945ms (Ohio) Network made differences!
- 36. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. PLM DB PoC result-2 1,442ms (Ireland) 2,339ms (Ohio) 3,063ms (Seoul) How we tested? 1. API servers in 3 regions called. 2. DynamoDB in Seoul for 1 hour. 3. DynamoDB returned 300 items(192KB). 4. Put data into ElastiCache Redis. 5. Get data from ElastiCache Redis 10x. 6. Logged the query + round trip time. 7. Measured average number. Why it took too long 1. DynamoDB worked well 2. ElastiCache Redis latency increased under high memory utilization and high traffic. What we learned? 1. Needs to keep some free memory space in operation for consistent low latency for Elasticache Redis. Network + ElastiCache Redis made differences!
- 37. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Key design is the key • Primary key selection for performance at scale – NoSQL itself is not a silver bullet • Understanding for Access Pattern for simplicity and cost - We have no secondary index
- 38. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. DynamoDB Schema Design for Bixby Partition1 Partition2 ….. PartitionN Partition1 Partition2 ….. PartitionN 3000 rows 1000 rows 1000 rows 1000 rows service_iduser_id xxx_iddomain Partition key Sort key 1000 rows 1000 rows 1000 rows domainuser_id xxx_idservice_id Partition key Sort key * Cardinality for IDs user_id(1) : service_id(1) : domain(N) : xxx_id(M)
- 39. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Choosing AWS database options for Bixby DynamoDB for Personal data Aurora MySQL : Member & admin ElastiCache Redis for caching data EMR for analytics S3 for object data EFS for low-latencyRDS MySQL : China RDS PostgreSQL : Geo functions
- 40. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. AWS Benefits ElastiCache Redis • Version : 2.8 - 5.0* • Cluster-mode* • Read scale-out with slaves* Aurora MySQL • Cloud based architecture : super performance with low-price • All instances are supported* • Read scale-out with up to 15 read replicas • Up to 5 cross region replication* • Custom endpoint* * Supported by AWS DynamoDB • Auto Scaling • 3 Availability Zones • VPC endpoint : Security • Consistent performance and low latency at any scale
- 41. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. Lessons learned • PoC • Database and API server have to be in the same place because of network • Needs to keep some free memory space in operation for consistent low latency for Elasticache Redis • DynamoDB table design • Making access pattern simple is the key • Spreading out traffic across all partitions is important • Less or no secondary key is better if possible • vs Bixby 1.0 • Working as one team easily through DB! • No license and operational cost with DynamoDB and EFS
- 42. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. 여러분의 피드백을 기다립니다! #AWSSummit 해시태그로 소셜 미디어에 여러분의 행사소감을 올려주세요. AWS Summit Seoul 2019 모바 일 앱과 QR코드를 통해 강연평 가 및 설문조사에 참여하시고 재미있는 기념품을 받아가세요 . 내년 Summit을 만들 여러분의 소중한 의견 부탁 드립니다. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 43. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved. 감사합니다! © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.