Use docker instead of rkt for regular etcdadm tasks (#4)

[ytsarev]

Motivation is to avoid serious memory leak on etcd nodes as in

• very high memory usage due to kernfs_node_cache slabs coreos/bugs#1927

Also it potentially mitigates issues #691 and #660 as the symptoms looks similar to our case.

[codecov-io]

Codecov Report

Merging #705 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #705   +/-   ##
=======================================
  Coverage   35.31%   35.31%           
=======================================
  Files          53       53           
  Lines        3047     3047           
=======================================
  Hits         1076     1076           
  Misses       1802     1802           
  Partials      169      169

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4637d39...062ea37. Read the comment docs.

[mumoshu]

@ytsarev Thank you very much for your effort on this 👍 I've read through/connected all the referenced links in mind and this change looks good to me.

[billyteves]

@mumoshu is the issue also exists in v0.9.6?

[mumoshu]

Yes, I believe so, unfortunately. 2017年6月21日(水) 18:25 billy ray teves <[email protected]>:

…

@mumoshu <https://github.com/mumoshu> is the issue also exists in v0.9.6? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#705 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AABV-feIAmUOMrv_DV-M6bmc4tawhXBoks5sGOGVgaJpZM4N8NX9> .

[Vince-Cercury]

If I select containerRuntime: docker in cluster.yml, is that supposed to use docker for etcd setup? or this cluster.yml option is only used for Workers? What about Controllers?

[mumoshu]

@Vincemd It is supposed to use docker only for containers managed by kubelets when containerRuntime: docker is set. Both controllers and workers runs kubelets for scheduling various pods, so controllers do get affected by the option.

On the other hand, etcd isn't managed by k8s/kubelets hence it is always run by rkt.