Experience lessons from architecture of zalo real time system
Download as PPTX, PDF•12 likes•5,889 views
Zalo is a messaging application with over 3 million users that exchanges 30 million messages per day. It uses a real-time system architecture with connection servers that allow for 1 million concurrent connections and 200,000 messages per second. The architecture is broken into independent layers including load balancing, web servers, searching, databases, and connection servers that use technologies like C/C++, Java, and event-based I/O. Lessons learned include avoiding HTTP long polling, optimizing for large not small data, and developing a custom memory allocator to avoid fragmentation issues.
Experience lessons from architecture of zalo real time system
- 1. Zalo Real-time System: Architecture & Experience By Nguyen Quang Nam Zing Senior Software Manager Web-Technical
- 2. Content • About Zalo • Technology in Zalo • Real-time System Architecture • Experience • Q&A
- 3. About Zalo
- 4. Features • Communication: message (voice, text), call • Social: “Nhật ký”, “Phòng trò chuyện”
- 6. Social
- 7. Updated stat • Users: 3.0M total • Messages: 30M per day
- 9. Technology • Programming languages: C/C++ and Java for services at server-side • Client platforms: IOS, Android, Symbian (S40, S60), Windows Phone (soon)
- 10. Technology (cont) Connection servers are key points of the Zalo real-time system Free / Opensource Inhouse Load balancing HA proxy ZaCacheProxy, ZaRouter Web servers Ngnix, Jetty ZaHttp Searching Elastic, Solr Prefix Search Service Database PostgreSQL, MySQL ZaDB, List32, List64, … Connection servers ZaCommServer, ZEventbus
- 12. N-Tier
- 13. ZaCommServer
- 14. ZaCommServer (cont) • Programming language: C/C • IO event mechanism: Epoll (libevent) • Binary protocol with authenticated connections & encrypted messages
- 15. ZaCommServer (cont) • Capacity per one box: • 1M concurrent connections tested • 200K throughput messages per second
- 16. Experience
- 17. Break the thing into multi-layers and let they be independent each other
- 18. Move biz out of high performance services
- 19. Lessons Zalo learned • Http long polling is not a good choice for a good app • Do not spend much time in optimizing small data • Memory fragment is a really big trouble in high performance services, develop our own memory allocator lib instead of using default system glibc
- 20. Q&A Contact: Nguyen Quang Nam [email protected] http://me.zing.vn/nam.nq