Optimizing Go for Clouds

on YouTube ↗01h 01m

" summary

Kubelet metric collection wastes CPU and GC time by allocating 5k DTO metrics on every scrape, producing ~8 MB and ~200k heap allocations per collect; goal is to cut allocations to about the raw data size. Reproduction uses Go microbenchmarks, pprof and benchstat to measure and compare. Fixes include caching DTOs, pre-allocating and reusing slices, and correcting label-sorting to avoid churn. Results drop allocations to <1MB / 25k allocations for 5k metrics and land within scrape timing.

AI-written summary · about

" chapters (63)

1. 0:00Introduction
2. 0:53Who am I
3. 1:54Why this topic
4. 3:57Outline
5. 4:20Theory
6. 4:49Speed Efficiency
7. 4:59Efficiency
8. 6:55Efficiency Speed
9. 7:17Counter Arguments
10. 7:54Why would you optimize
11. 10:12Now hearing all of this
12. 10:26Optimization kills readability
13. 11:44Premature optimization
14. 12:02Hardware is getting faster
15. 14:37Scale horizontally
16. 14:47Multithreaded
17. 15:18Time to market
18. 15:41Go is efficient
19. 17:12Datadriven decisions
20. 17:25Optimization challenges
21. 19:55Open source issue on Kubernetes
22. 22:11Optimizing Kubernetes
23. 26:12Test Driven Development
24. 27:05Benchmark Driven Optimization
25. 28:15Macro Benchmarks
26. 29:44Benchmark Tool
27. 32:12Test Data
28. 33:40Compilation
29. 34:40Global Variable
30. 35:14Benchmarks
31. 36:02Understanding the results
32. 39:10Finding bottlenecks
33. 39:54Pprov
34. 41:14Malloc
35. 41:31Garbage Collection
36. 41:52Share Polar Signals
37. 42:19Memory
38. 43:08Magic Graph
39. 43:35Prometheus Process Metric
40. 43:57Peak
41. 44:16Collect
42. 45:44Algorithm Level Optimization
43. 47:10Complexity
44. 47:44Cache
45. 48:52Race
46. 49:22Single Flight
47. 50:16Implementation
48. 50:36Benchstat
49. 51:03Collect Container Info
50. 51:28String Slice
51. 51:43Array Slice
52. 53:31Preallocate
53. 55:06Magic values
54. 55:41Results
55. 56:02Add Metric
56. 56:29Debug
57. 56:40C Advisor
58. 56:55Client Golang
59. 57:39Finally
60. 58:11Documentation
61. 59:01Pattern
62. 59:23Summary
63. 59:38Learning

" related talks

• 19mHeap Optimization for Go SystemsNishant Roy · SREcon20 Americas
• 51mObscure Go OptimisationsBryan Boreham · GopherCon UK 2022
• 14mOptimizing Go code without a blindfoldDaniel Martí · dotGo 2019
• 40mGolang memory footprintLiviu Esanu · code.talks 2019
• 46mSimplifying Scalable Cloud Application Development with GoMandy Waite · Code Mesh 2015
• 48mGolang's GarbageAndrey Sibiryov · SREcon17 Asia/Australia
• 51mEvolving the Go Memory Manager's RAM and CPU EfficiencyMichael Knyszek · GopherCon 2020
• 56mThe Key to Go Efficiency is Just a Few Go Runtime Metrics Away!Bartłomiej Płotka, Arianna Vespri · GopherCon UK 2024