Scaling Coder
Coder's control plane (coderd
) and workspaces are deployed in a Kubernetes
namespace. This document outlines vertical and horizontal scaling techniques to
ensure the coderd
pods can accommodate user and workspace load on a Coder
deployment.
Vertical scaling is preferred over horizontal scaling!
Vertical Scaling
Vertical scaling or scaling up the Coder control plane (which consists of the
coderd
pod and any additional replicas) is done by adding additional computing
resources to the coderd
pods in the Helm chart's values.yaml
file.
Download the values file for a deployed Coder release with the following command:
helm get values coder > values.yaml -n coder
Experiment with increasing CPU and memory requests and limits as the number of
workspaces in your Coder deployment increase. Pay particular attention to
whether users have their workspaces configured to auto-start at the same time
each day, which produces spike loads on the coderd
pods. To best prevent Out
of Memory conditions aka OOM Kills, configure the memory requests and limits to
be the same Gi values. e.g., 8Gi
Increasing
coderd
CPU and memory resources requires sufficient Kubernetes node machine types to accomodatecoderd
, Coder workspaces and additional system and 3rd party pods on the same cluster namespace.
These are example values.yaml
resources for coderd
's CPU and memory for a
larger deployment with hundreds of workspaces autostarting at the same time each
day:
coderd:
resources:
requests:
cpu: "4"
memory: "8Gi"
limits:
cpu: "8"
memory: "8Gi"
Leading indicators of undersized coderd
pods include users experiencing
disconnects in the web terminal, a web IDE like code-server or slowness within
the Coder UI dashboard. One condition that may be occuring is an OOM Kill where
one or more coderd
pod fails, restarts or fails to restart and enteres a
CrashLoopBackOff status. If coderd
restarts and there are active workspaces
and user sessions, they will be reconnected to a new coderd
pod causing a
disconnect situation. As a Kubernetes administrator, you can also notice
restarts by noticing frequently changing and low AGE
column when getting the
pods:
kubectl get pods -n coder | grep coderd
Horizontal Scaling
Another way to distribute user and workspace load on a Coder deployment is to
add additional coderd
pods.
coderd:
replicas: 3
Coder load balances user and workspace requests across the coderd
replicas
ensuring sufficient resources and response time.
There is not a linear relationship between nodes and
coderd
replicas so experiment with incrementing replicas as you increase nodes. e.g., 8 nodes and 3coderd
replicas.
Horizontal Pod Autoscaling
We do not recommend using Horizontal Pod Autoscaling with Coder. In some environments, this will lead to errors with workspace builds.