Increase default node cpu size from 1 to 4. #335
Conversation
Each Revision needs most of a CPU. Conformance tests need two Revisions. And we have all the control plane stuff running. So we need bigger nodes.
google-prow-robot
added
the
size/XS
|
#291 (comment) has some more context on why this change is necessary. |
|
What do you think about adding some docs at https://github.com/elafros/elafros#latest-release about minimum cluster requirements as well? |
|
I created a new cluster with n1-standard-4 nodes, and conformance tests now consistently pass (on this branch). 🎉 |
| @@ -29,6 +29,7 @@ To use a k8s cluster running in GKE: | |||
| --cluster-version=1.9.2-gke.1 \ | |||
| --zone=us-east1-d \ | |||
| --scopes=cloud-platform \ | |||
| --machine-type=n1-standard-4 \ | |||
There was a problem hiding this comment.
Since this will be significantly more expensive for users, can we add a comment explaining why this machine type is necessary?
There was a problem hiding this comment.
I added a comment below explaining why this is recommended.
|
/retest |
| @@ -29,6 +29,7 @@ To use a k8s cluster running in GKE: | |||
| --cluster-version=1.9.2-gke.1 \ | |||
| --zone=us-east1-d \ | |||
| --scopes=cloud-platform \ | |||
| --machine-type=n1-standard-4 \ | |||
There was a problem hiding this comment.
I added a comment below explaining why this is recommended.
|
I'd like to get this merged now so we can be in a state where folks can follow the setup instructions and have passing conformance tests, but @josephburnett could you follow up with some docs or an issue to create docs about minimum system requirements for folks that want to use Elafros? |
|
@bobcatfish. Ack. I'll add those docs. |
Previously we were requesting 2.5% of a CPU for Revisions which is not enough to run any meaningful load. And the scheduler will happily pack way too many Pods into each node and therefore the cluster autoscaler won't kick in when needed.
The Elafros autoscaler commit (#229) increases the CPU request per Pod to 1 full CPU. This was necessary to make autoscaling work correctly. However the default, tiny cluster of 3, 1-core nodes simply couldn't fit the Elafros control plane (Istio, Controllers, Webhooks) and two Revisions. Even 2-core machines are a little small.
So I'm increasing the default to 4 cores and autoscaling will bring the cluster down to 1 node when not used.