フェアリーデバイセズがRustを 採用して3年が過ぎました

Rust.Tokyo 2022 スポンサーセッション

Transcript

1. 2022-09-23
 
 フェアリーデバイセズ株式会社
 Fairy Devices Inc.
 1 Rust.Tokyo 2022 スポンサーセッション
 
 フェアリーデバイセズがRustを
 採用して3年が過ぎました

2. About me
 • Twitter/GitHub: @emergent
 • Since 2018.9
 ◦ Manager of Product Development Dept.
 • Publication
 ◦ 『実践Rustプログラミング入門』共著
 ◦ WEB+DB PRESS No.129 (interview)
 ◦ WEB+DB PRESS No.131 (plan)
 ◦ 技術書典『機械学習の炊いたん』
 2 Satoshi Yoshikawa / 吉川哲史
5. 5 Company history

6. 6 Company history

7. Products @2018-2019
 Audio libraries written in C++ WebSocket server in C++ and the others in Python and TypeScript 7 Tumbler
8. Difficulties with services handling audio
 • Theoretical and implementation difficulties of audio signal processing, speech recognition, etc.
 • Difficulties in APIs as a pathway for audio data
 8
9. Difficulties with services handling audio
 • Theoretical and implementation difficulties of audio signal processing, speech recognition, etc.
 • Difficulties in APIs as a pathway for audio data
 9 Asynchronous communication of audio stream

10. Handling audio streams
 10 audio application speech recognition Audio stream Client
 Server
 Text

11. Handling audio streams
 11 audio application speech recognition Audio stream Client
 Server
 Text
 Asynchronous

12. Products @2018-2019
 Audio libraries written in C++ WebSocket server in C++ and the others in Python and TypeScript 12 Tumbler
13. Hard to maintain with C++...
 13 Write once in C++, run forever. PIEN CEO / CTO
 Masato Fujino
 ※He didn’t say so, (actually). 

14. Is there the alternative of C++?
 • Want to increase software quality
 ◦ No more segmentation fault
 • but not at the expense of performance
 
 14
15. There is Rust
 15 https://www.rust-lang.org/
16. Road to use Rust to our products
 1. Write codes in Rust
 2. Write codes in Rust
 3. Write codes in Rust
 4. Say “Rust is good!” to collegues
 5. Say “Rust is good!” to CEO
 6. Say “Rust is good!” to CEO
 7. Say “Rust is good!” to CEO
 8. Write codes in Rust
 9. Write codes in Rust
 10. Write codes in Rust
 16
17. Steps with Rust
 17 Small start
 (microservice)
 Replacement
 New product
 Challenges
 
 ・HTTP server
 ・Connecting Redis
 Challenges
 
 ・WebSocket server 
 ・Connection control 
 Challenges
 
 ・HTTP server
 ・WebSocket API
 ・RDBMS
 ・Lambda function
 ・Layered architecture 

18. Products with Rust
 2019 2020 2021 2022 CATS Backend microservice (small start) Nyaprus Speech Recognition API (replacement) LINKLET (new product) Product development
 Personal activity
 WEB+DB PRESS No.131 Publish Interviewed 18 …and many trivial works in Rust
 Publish(plan)
19. Products with Rust
 2019 2020 2021 2022 CATS Backend microservice (small start) Nyaprus Speech Recognition API (replacement) LINKLET (new product) Product development
 Personal activity
 WEB+DB PRESS No.131 Publish Interviewed 19 …and many trivial works in Rust
 Publish(plan)
20. CATS
 Purpose : to throttle simultaneous connections of APIs
 • Concurrent Access Throttling System
 • A simple HTTP server with Redis backend
 20 Application using Actix web Cloud Memorystore (on GCP) HTTP request HTTP response ≒ Redis

21. Architecture
 21 CATS API server 1 API server 2 API server 3 Engine 1 Engine 2 Engine i … … Frontend Backend (speech recognition) CATS
22. Wrote many “Future” functions
 22 … We use async/await since 1.39.0 
 CATS
23. Good points
 • Operates stably than we thought
 • High performance and high load capacity
 
 • …And is contributing our business!
 23 CATS
24. First product in Rust 🎉
 24 CATS
25. Side stories
 • Actix Web: soon to be version 1.0 or not?
 • Participated in Rust communities and gatherings.
 ◦ “Is anyone familiar with Future anywhere?”
 25 CATS
26. Products with Rust
 2019 2020 2021 2022 CATS Backend microservice (small start) Nyaprus Speech Recognition API (replacement) LINKLET (new product) Product development
 Personal activity
 WEB+DB PRESS No.131 Publish Interviewed 26 …and many trivial works in Rust
 Publish(plan)
27. Nyaprus - Speech Recognition API
 Purpose: add functionality to speech recognition API
 • A frontend server of speech recognition engine
 • Has WebSocket API
 • Using warp 27 I want to use tokio-tungstenite
28. • Replaced from C++ and Java
 • Implemented
 ◦ a WebSocket server
 ◦ a proxy server having connection pool
 Implementation
 28 Nyaprus Before
 After
 WS server in C++ SR frontend in Java (proprietary) SR engine (proprietary) WS server in Rust Proxy server in Rust SR engine (proprietary) WS: WebSocket 
 SR: Speech Recognition 

29. Techniques on Rust
 • trait Stream / Sink ( on WebSocket server )
 • Connection Pooling ( on proxy server )
 29
30. Stream / Sink
 30 Nyaprus Client
 Nyaprus
 Speech
 Recognition
 Engine
 rx tx sink stream
31. Stream / Sink
 31 Nyaprus Stream
 Stream
 Sink
 Sink

32. Stream / Sink
 32 Nyaprus …and join
 Client
 Nyaprus
 Speech
 Recognition
 Engine
 rx tx sink stream
33. Connection pooling
 Problem
 Recognition speed decreases when disconnected
 due to cache discard in the speech recognition engine
 Solution
 Keep and reuse connections in connection pools
 33 Nyaprus
34. Connection pooling with bb8
 34 Nyaprus ……
 ……
 ……
 Implemented trait

35. Connection pooling with bb8
 35 Nyaprus Set up

36. Proxy - using tokio::io::copy
 36 Nyaprus
37. Good points
 • First replacement 🎉
 • To improve development productivity
 ◦ Used task runner (cargo-make)
 ◦ Used container (Docker, CI/CD)
 ◦ Created testing tool (in Rust)
 37 Nyaprus
38. Side stories
 • Cache problems were discovered just before release
 ◦ Prototype implemented in 1 day
 ◦ After 2 weeks of testing & brushing up
 ◦ Managed to release it! 🎉
 38 Nyaprus
39. Products with Rust
 2019 2020 2021 2022 CATS Backend microservice (small start) Nyaprus Speech Recognition API (replacement) LINKLET (new product) Product development
 Personal activity
 WEB+DB PRESS No.131 Publish Interviewed 39 …and many trivial works in Rust
 Publish(plan)
40. LINKLET (backend)
 • API backend of LINKLET https://linklet.ai/
 40
43. Challenges
 • “Web service” development (for me)
 • Developed by team with multiple people
 • Continuous development and release
 43 LINKLET
44. Architecture (part)
 44 api-server Web Frontend notification zoom auth0 tlpf-connector mqtt-connector db-connector Auth0 Zoom Device Controller RDBMS httpclient WebSocket notificaiton HTTP(S) Webhook (HTTP(S)) cipher LINKLET
45. Multiple crates in one workspace
 45 LINKLET
46. Challenges
 • Asynchronous communication with LINKLET devices
 ◦ Taking care of the state of devices
 ▪ Power ON / OFF
 ▪ Online / Offline
 ◦ Non-blocking API
 ▪ HTTP and WebSocket (push)
 46 LINKLET
47. Challenges (with Rust)
 • Wrestling with Diesel (and SQL)
 • Layered architecture (in progress)
 ◦ Using mock (mockall crate)
 • AWS Lambda function in Rust
 47 LINKLET
48. Advantages in Rust
 • Refactoring with peace of mind😌
 ◦ by
 ▪ Compilation errors are well attended
 ▪ Interfaces are well typed
 48 LINKLET
49. Steps with Rust
 49 Small start
 (microservice)
 Replacement
 New product
 Challenges
 
 ・HTTP server
 ・Connecting Redis
 Challenges
 
 ・WebSocket server 
 ・Connection control 
 Challenges
 
 ・HTTP server
 ・WebSocket API
 ・RDBMS
 ・Lambda function
 ・Layered architecture 

50. Conclusions
 • Rust is powerful to create web applications
 ◦ But still less than 30% of our total products
 50
51. Futures
 • Grow our products more with Rust (and modern techs)
 ◦ more reliable
 ◦ more speedy
 ◦ higher performance
 51
52. We are hiring!
 • Frontend / Backend developer
 • Device Engineer
 • Product Manager
 • …and more
 52
53. We are hiring!
 • Frontend / Backend developer
 • Device Engineer
 • Product Manager
 • …and more
 53 Please check the job board out!

54. Thank you!
 54