This message was deleted.
# troubleshootings
Slackbot
06/21/2023, 3:26 PMThis message was deleted.
c
churro
06/22/2023, 2:31 AMif you are in k8s i recommend using mm-less. each task gets its own k8s pod, let cgroups handle resource isolation
j
Julian Reyes
06/22/2023, 9:38 AMwe have not used that yet, thought it was experimental
c
churro
06/29/2023, 6:27 PMwe use it in production for a lot of our clusters…it works just fine. Don’t be scared to try it 🙂
j
Julian Reyes
06/29/2023, 6:29 PMMight try it first in the dev cluster. Is there a big shift in configuration with regards to helm charts and middleManagers?
c
churro
06/29/2023, 8:30 PMput the values in the overlords and dont deploy the middle managers….
j
Julian Reyes
06/29/2023, 9:51 PMIn my case that would be in the coordinators pods then
Julian Reyes
06/29/2023, 10:14 PMAnd I guess the nodes for coordinators would need to be bigger? As well as replicas.
I wonder if that config reduces cost in resources
Julian Reyes
06/29/2023, 10:19 PMIf you are using helm templates or k8s yamls, would you mind sharing an example?
c
churro
06/30/2023, 10:53 PMi use the operator, so i can give you the configs on monday. So i would have independent overlord / coordinator nodes, keep your coordinator about the same in terms of RAM / CPU. The tasks get launched as k8s pods, thus you don’t need to statically allocate resources up front, whatever your eks cluster has basically. If you need more resources, add nodes, if you are not using enough, remove them. We use it with autoscaling here and eks…which has worked out quite nicely in terms of a decent cost savings.
churro
06/30/2023, 10:53 PMill share the overlord config, basically don’t deploy the MM in your helm chart, and add some config values to the overlord node and add the extension.
j
Julian Reyes
07/01/2023, 12:59 PMWhat kind of adapter do you use ? I posted this in the k8s channel to try to get some advice
https://apachedruidworkspace.slack.com/archives/C04Q0047B4M/p1688144035521209