production: an owner's manual

from exec(ut) 2018

Transcript

1. production: an owner's manual
2. hello!
3. broken computers
5. getting sidetracked now so sorry* * not sorry
9. back to serious business
10. !
12. a production system is a system that serves real users
13. the goal of operations is to ensure services are reliable
14. in order to provide a good user experience
16. failure
17. app
18. app linux kernel cpu dram disk network power supply switches load balancer dns submarine cables routers fiber
19. app linux kernel the cloud
21. • cosmic rays • disk failure • power outages • software bugs • ...
22. entropy
24. capacity
28. cascading failure
30. system design
31. redundancy
32. "
33. scale
35. "
36. p1 m3 c1 m2 m1 p2 c2
37. data storage
38. "
39. "
40. protocols
42. monitoring
43. many components many req/s
45. measure all the things?
46. ✅ ⏱
47. golden signals • latency • traffic • errors • saturation
48. golden signals • latency • traffic • errors • saturation
49. golden signals • latency • traffic • errors • saturation
50. golden signals • latency • traffic • errors • saturation
51. golden signals • latency • traffic • errors • saturation 0 - 50 [1620]: ∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎ (74.55%) 50 - 100 [ 447]: ∎∎∎∎∎∎∎∎∎∎ (20.57%) 100 - 150 [ 49]: ∎ (2.25%) 150 - 200 [ 15]: (0.69%) 200 - 250 [ 15]: (0.69%) 250 - 300 [ 10]: (0.46%) 300 - 350 [ 6]: (0.28%) 350 - 400 [ 1]: (0.05%) 400 - 450 [ 0]: (0.00%) 450 - 500 [ 4]: (0.18%)
52. golden signals • latency • traffic • errors • saturation
53. saturation traffic latency errors
55. humans
57. oops, deleted the database
58. bad human!
59. why does this button even exist?
60. app linux kernel cpu dram disk network power supply switches load balancer dns submarine cables routers fiber
61. app linux kernel cpu dram disk network power supply switches load balancer dns submarine cables routers fiber humans
62. app linux kernel cpu dram disk network power supply switches load balancer dns submarine cables routers fiber humans h u m a n s
63. epic failure is almost always systemic
64. failure
65. recap
66. • a production system serves real users • users like things that work and are fast • epic failure is almost always systemic
67. thx @igorwhilefalse