Swifty microservices in production

Transcript

1. Swifty microservices in production Marcin Kliks Twitter: vi4m http://www.marcinkliks.pl
2. About • Marcin Kliks • Devops Teams Lead @ Allegro • Ralph CMDB http://github.com/allegro/ ralph • Loves clean & performant code
4. http://www.allegro.pl
5. Backend stack Website (business logic) PaaS • 500+ microservices 2 Data Centers
6. The story of SSS
9. 110 KB
10. 3s Time budget Many reasons, images was one of them
11. Smaller images • WebP? • PNG and JPEG - most compatible ones • Performant solution for thousands of images
13. Let’s shrink it!
14. Let’s shrink it! 30-40%
15. Mozjpeg 30%
16. Recompression is easy, right?
18. Image Compression • Allocate memory • Decompress • Recompress (quantization) • Free memory 40 x RAM 4 MB image -> 160 MB RAM
19. Lots of images
20. Used Allocated Offer pool Mesos cluster
21. Used Allocated … Offer pool Mesos cluster
22. C library integration • How easy is it to integrate and to compile? • How can you debug it? • What’s the performance / latency cost? • How do you release it? • Types conversion?
23. C library integration Single compiler Debugging Perf. Penalty Types briding Golang ✅ (CGo) No 5x ❌ Java ❌ No 10 - 25x ❌ Python ❌ No Significant ❌ Swift ✅ Yes, Xcode Zero, native calls ✅
24. C library integration Single compiler Debugging Perf. Penalty Types briding Golang ✅ (CGo) No 5x ❌ Java ❌ No 10 - 25x ❌ Python ❌ No Significant ❌ Swift ✅ Yes, Xcode Zero, native calls ✅
25. Swift c interop
26. Swift c interop String -> const char * Implicit type conversion
27. from: raywenderlich.com
28. Performance • 3x lower memory usage than Java • Value types! • No Garbage collector! • Predictive performance
29. How?
31. Swift
32. Swift Clang / LLVM
33. Swift Clang / LLVM
34. New language, new world Research project
35. Used components • Swift 4.1 • SPM • Vapor 2 (porting to 3) • GCD • HTTPS / TCP Client
36. Preparations • Choosing Server Side Swift framework • Prepare runtime environment on Mesos • In-house Docker images for Swift • Plug into ecosystem • CI • One click deployment • Security checks • Compliance checks • Autodiscovery, logs, metrics … • Implement Microservice Contract
37. Microservice contract 1
38. Implementing microservice with Swift /status/ping addMicroserviceContractEndpoints(route)
39. Implementing microservice with Swift Configuration Metrics Logs
40. Implementing microservice with Swift Configuration Metrics Logs /status/ping
41. Implementing microservice with Swift Vault Graphite + Grafana Logstash + Kibana /status/ping
42. Implementing microservice with Swift getConfigurationVariable(„DB_PASS”) Vault Graphite + Grafana Logstash + Kibana /status/ping
43. Vault configuration-file-test.yml database-name: [name]
44. Implementing microservice with Swift Vault Graphite + Grafana Logstash + Kibana logger.info("Startup") /status/ping
45. Kibana logs
46. Implementing microservice with Swift Configuration Service pushMetrics(metricName:value) Graphite + Grafana Logstash + Kibana /status/ping
47. Grafana
48. Implementing microservice with Swift • Helper library („swidamio”) • Configuration Service Client • Graphite client • Logging client • Microservice Contract
49. Integrating C libraries 2
50. System or vendoring? • Linking to system libraries (installed with homebrew, apt) • Faster compiles • Package manager responsible for updates • Sources included(„vendored”) within the app • Customizable • Longer compiles
51. System or vendoring? • `brew install mozjpeg`
59. Did we use pkg-config ?
60. pkg-config
61. pkg-config
66. custom-settings.xcconfig
68. Debugging both 
 C + Swift code in XCode
69. Challenges
70. Segmentation fault
71. Challenges • Catching fatal errors in production / Sentry • All functions throwable • Memory profiling in production • C-based profilers?
72. Deploy 3
73. 0-dependencies deploy 1. Compilation on Docker 2. package.zip with all libs 3. Unzip & run native on mesos
74. Demo
76. Results R
77. 170 KB 90 KB
78. Swift minimal 25 MB of RAM Vapor 2.
79. 250 MB of RAM Java, springboot
80. Swift - during compression of images Limit ~ 49 MB of RAM
81. Startup time 0,2 sec ~ Servless like ~
83. Does it scale? Java 
 10 instances Swift 
 100 instances
84. 2 * 100 instances in 40 seconds
 - including health checks
 - nothing preinstalled on clusters
85. What’s next? • We just got started (PNG, one Showcase Area) • Next step: compress more areas with PNG + JPG
86. Quanta Quanta - Open Source in November! http://allegro.tech
87. Good parts • FAST, predictive performance, and fast to scale • Amazing c-libraries integration • Powerful language with safe code constructs
88. Missing parts… • Framework agnostic libraries like Logging, Metrics • Boundles • Stacktraces for fatal C / Unwrap crashes • Memory profiling on Linux • Better tooling
89. What is Swift good for?
90. The best for • Performance-oriented microservices • Safer C libraries wrappers • ML / CPython integration (Dynamic Callable) • …
91. Thank you Marcin Kliks marcin.kliks@icloud.com @vi4m Q & A