Videos / DevOps

Episode 9 of 14 · DevOps, end to end

Kubernetes, part two: scheduling, storage, and what breaks at three in the morning

Chapters

1. The label on the box0:00
2. A request is a reservation0:16
3. Cpu is compressible0:32
4. There is no slower byte0:51
5. Who gets evicted first1:08
6. What to actually set1:26
7. Filter, then score1:44
8. The bluntest lever2:01
9. Affinity has grammar2:16
10. The node refusing2:36
11. Skew across the zones2:54
12. Placement is a decision3:14
13. The board gets wiped3:33
14. The claim and the volume3:50
15. The class that provisions4:08
16. One node at a time4:28
17. Identity that comes back4:45
18. What it does not do5:02
19. Three kinds of door5:21
20. One balancer, many routes5:40
21. TLS stops at the edge5:58
22. The annotation problem6:17
23. Three objects, three owners6:37
24. Routing in the spec6:56
25. Three separate dials7:14
26. The arithmetic7:34
27. Scale on the queue7:53
28. The vertical one8:11
29. One pending pod8:31
30. Larger than any machine8:49
31. One word, one layer9:07
32. Never scheduled at all9:24
33. It never got the image9:42
34. Started, then exited9:57
35. The kernel decided10:16
36. The probe that lies10:34
37. The whole thing10:52

Kubernetes, part two: scheduling, storage, and what breaks at three in the morning. A node has a fixed amount of room, and the scheduler packs pods onto it by reading a label on the box, never by weighing the box. That label is two numbers per container. Requests,

Mark as watchedSaved in this browser only.