Closed jmcarp closed 7 years ago
All minions are currently c3.2xlarge
which provides 8 VCPUs, and 15GB of RAM; none of them have 12GB free to handle this workload so we need to resize.
Our options are:
r4.2xlarge
raises memory to 60GB
r4.xlarge
raises memory to 30GB, but halves the VCPUs to 4
I think we should switch to r4.xlarge
as none of our minions have used more than 4% CPU in the last two weeks.
Sounds good, I'll send in that change. We could also use taint / toleration to schedule elastic to high-memory boxes later on, but no need for that now.
Agreed, there's an issue to do that here: https://github.com/18F/kubernetes-broker/issues/24 which maybe we'll do when making HA services in the next PI?
As @kynetiv pointed out, the 6x plan on the legacy broker allocated 6gb of heap, since heap and total memory were set to the same value. Which means that the 6x plan here has half the heap of the 6x legacy plan. This patch adds a 12x plan with 6gb of heap, which should correspond better to the legacy 6x plan. I also bumped the persistent disk allocation to 20gb.
I'm wondering if we'll want more and/or bigger boxes in the cluster to support this plan--I'll look into this, and hopefully get an opinion from @cnelson too. Then we'll spin up an instance for college scorecard and see if performance is adequate.