From eval to prod: How a Service Mesh Helped Us Build Production Cloud-Native Services

In this talk, Israel Sotomayor (Infrastructure Engineer) covers how Moltin used Linkerd in production from day 0; in other words, this is a look at going from evaluation through running in production.

Israel explains how Moltin systems started as a monolithic beast that needed a full architectural revamp to decompose it into microservices. In shifting to microservices, several key concerns presented themselves and his team was able to use Linkerd throughout the process to solve practical problems.

Israel shares his insights, lessons learned, and practical tips you can use to make adopting cloud-native architectures achievable using a service mesh.

https://www.youtube.com/watch?v=VoYAtCVz3Ig

Transcript

1. From eval to prod How a service mesh helped us build production cloud-native services @zot24 ! from @moltin "
2. What will you learn today? ! You'll be able to understand the needs and problems we faced when we built our distributed system. @zot24 !
3. Summary ! 4 What we do at moltin? 4 State of art before (micro)services 4 State of art after (micro)services 4 Microservices 4 Service Mesh 4 What's next? @zot24 !
4. What we do at moltin? ! 4 API-First 4 Commerce experience @zot24 !
5. State of art before (micro)services ⾢
6. 2 years ago ⌛ 4 A huge monolithic app 4 Highly coupled code 4 No tests at all 4 Poor performance 4 High latency responses @zot24 !
7. =
8. !
9. !
10. So this is how we looked like at that point
11. @zot24 !
12. Our first tasks ! 4 Increase performance ! 4 Decrease response latency " @zot24 !
13. How did we solve it? 4 Identified the most use parts of our system 4 Identified the slowest one 4 Improve it! @zot24 !
14. The most used and slowest performing part 4 Authentication system ! 4 Around ~500ms to request an access token " @zot24 !
15. How did we approach a solution? 4 Tweaking ! 4 From 500ms to 250ms = " 4 Decoupling # 4 Our first "(micro)service" $ 4 The right tool % for the job & 4 From 250ms to ~40ms ' = ( @zot24 !
16. So this is how we looked like at that point
17. @zot24 !
18. State of art after (micro)services ▶ Microservices
19. Why (micro)services? 4 We needed to improve our system performance ! and reduce our request latencies " @zot24 !
20. ! but... how?
21. @zot24 !
22. !
23. ! How did we approach it?
24. ! !"#☝
25. !
27. @zot24 !
28. So this is how we looked at that point
29. @zot24 !
30. State of art after (micro)services ▶ Service Mesh
31. What's a Service Mesh? A dedicated infrastructure layer for service to service communication decoupled from your application code and focus on services and requests @zot24 !
32. Why did we need a service mesh in the first place?
33. Service to service communication needs to be managed, monitored and controlled
34. ! Observability
35. ! Development 4 Did my request got from A to B? 4 Where is it hanging? 4 Why my system is so slow? @zot24 !
36. ! Production 4 Improve resilience - Circuit breaking 4 Improve success rate - Retries 4 Reduce tail latency - Load Balancing 4 Metrics and Tracing - Observability @zot24 !
37. Circuit Breaking problem @zot24 !
38. Latency aware as a problem @zot24 !
39. What's out there?
40. Service Mesh Solutions 4 Linkerd 4 Conduit (Linkerd for k8s) 4 Envoy (from Lyft) 4 Istio @zot24 !
41. Why Linkerd? 4 Vendor agnostic 4 Stronger option 4 Well adopted by the community 4 Trusted by multiple companies in production 4 Great support @zot24 !
42. What is it really? Is just another proxy but the feature set and focus is very different @zot24 !
43. A -> B
44. A -> L -> B
45. Will be my system's performance affected by introducing Linkerd?
46. YES it will!
47. Your best case will take a hit (a small performance implication)
48. Your worst case will have a huge performance improvement
49. What's your goal in a distributed system? 4 Reduce your tail latencies p99 4 Have a stable and resilient system @zot24 !
50. What else can we do with Linkerd?
51. Load Balancing Expirations Logging Circuit Breaking Error tracking Canary Releases Tracing Retries Metrics Back-offs Retry budgets Dynamic routing Service Discovery Timeouts
52. How much of all this are we using?
53. ! Not even the 15% of all features
54. !
55. ! it just worked out of the box " with minimal configuration
56. So this is how we looked at that point
57. @zot24 !
58. What's next?
59. We want to do more cool ! things 4 Canary deployment 4 Shadowing traffic 4 Dark Launches @zot24 !
60. Canary deployment @zot24 !
61. Canary deployment @zot24 !
62. Canary deployment @zot24 !
63. Canary deployment @zot24 !
64. Canary deployment @zot24 !
65. Canary deployment @zot24 !
66. Shadowing traffic @zot24 !
67. Dark Launches @zot24 !
68. Looking forward to...? 4 Adopt to Kubernetes 4 Eval Conduit @zot24 !
69. Take Away ! 4 Network communication is complicated 4 Apps shouldn't have to reinvent this wheel 4 Libraries are language specific @zot24 !
70. so...
71. You will need a service mesh if you are building a distributed system
72. demo⌛❓
73. Thanks you! !" 4 I'm Israel Sotomayor !" 4 Infrastructure Engineer # 4 Digital Nomad $ 4 Enjoy automating things % 4 ❤ Traveling ✈ @zot24 ! from @moltin "