coder/docs/admin/scale.md

We scale-test Coder with the [same utility](#scaletest-utility) that can be used in your environment for insights into how Coder scales with your infrastructure.

## General concepts

Coder runs workspace operations in a queue. The number of concurrent builds will be limited to the number of provisioner daemons across all coderd replicas.

- **coderd**: Coder’s primary service. Learn more about [Coder’s architecture](../about/architecture.md)
- **coderd replicas**: Replicas (often via Kubernetes) for high availability, this is an [enterprise feature](../enterprise.md)
- **concurrent workspace builds**: Workspace operations (e.g. create/stop/delete/apply) across all users
- **concurrent connections**: Any connection to a workspace (e.g. SSH, web terminal, `coder_app`)
- **provisioner daemons**: Coder runs one workspace build per provisioner daemon. One coderd replica can host many daemons
- **scaletest**: Our scale-testing utility, built into the `coder` command line.

```text
2 coderd replicas * 30 provisioner daemons = 60 max concurrent workspace builds
```

## Infrastructure recommendations

### Concurrent workspace builds

Workspace builds are CPU-intensive, as it relies on Terraform. Various [Terraform providers](https://registry.terraform.io/browse/providers) have different resource requirements. When tested with our [kubernetes](https://github.com/coder/coder/tree/main/examples/templates/kubernetes) template, `coderd` will consume roughly 8 cores per 30 concurrent workspace builds. For effective provisioning, our helm chart prefers to schedule [one coderd replica per-node](https://github.com/coder/coder/blob/main/helm/values.yaml#L110-L121).

To support 120 concurrent workspace builds, for example:

- Create a cluster/nodepool with 4 nodes, 8-core each (AWS: `t3.2xlarge` GCP: `e2-highcpu-8`)
- Run coderd with 4 replicas, 30 provisioner daemons each. (`CODER_PROVISIONER_DAEMONS=30`)
- Ensure Coder's [PostgreSQL server](./configure.md#postgresql-database) can use up to 2 cores and 4 GB RAM

## Recent scale tests

| Environment        | Users | Concurrent builds | Concurrent connections (Terminal/SSH) | Coder Version | Last tested  |
| ------------------ | ----- | ----------------- | ------------------------------------- | ------------- | ------------ |
| Kubernetes (GKE)   | 1200  | 120               | 10,000                                | `v0.14.2`     | Jan 10, 2022 |
| Docker (Single VM) | 500   | 50                | 10,000                                | `v0.13.4`     | Dec 20, 2022 |

## Scale testing utility

Since Coder's performance is highly dependent on the templates and workflows you support, we recommend using our scale testing utility against your own environments.

The following command will run our scale test against your own Coder deployment. You can also specify a template name and any parameter values.

```sh
coder scaletest create-workspaces \
    --count 1000 \
    --template "kubernetes" \
    --concurrency 0 \
    --cleanup-concurrency 0 \
    --parameter "home_disk_size=10" \
    --run-command "sleep 2 && echo hello"

# Run `coder scaletest create-workspaces --help` for all usage
```

> To avoid potential outages and orphaned resources, we recommend running scale tests on a secondary "staging" environment.

The test does the following:

1. create `1000` workspaces
1. establish SSH connection to each workspace
1. run `sleep 3 && echo hello` on each workspace via the web terminal
1. close connections, attempt to delete all workspaces
1. return results (e.g. `998 succeeded, 2 failed to connect`)

Concurrency is configurable. `concurrency 0` means the scaletest test will attempt to create & connect to all workspaces immediately.

## Autoscaling

We generally do not recommend using an autoscaler that modifies the number of coderd replicas. In particular, scale
down events can cause interruptions for a large number of users.

Coderd is different from a simple request-response HTTP service in that it services long-lived connections whenever it
proxies HTTP applications like IDEs or terminals that rely on websockets, or when it relays tunneled connections to
workspaces. Loss of a coderd replica will drop these long-lived connections and interrupt users. For example, if you
have 4 coderd replicas behind a load balancer, and an autoscaler decides to reduce it to 3, roughly 25% of the
connections will drop. An even larger proportion of users could be affected if they use applications that use more
than one websocket.

The severity of the interruption varies by application. Coder's web terminal, for example, will reconnect to the same
session and continue. So, this should not be interpreted as saying coderd replicas should never be taken down for any
reason.

We recommend you plan to run enough coderd replicas to comfortably meet your weekly high-water-mark load, and monitor
coderd peak CPU & memory utilization over the long term, reevaluating periodically. When scaling down (or performing
upgrades), schedule these outside normal working hours to minimize user interruptions.

### A note for Kubernetes users

When running on Kubernetes on cloud infrastructure (i.e. not bare metal), many operators choose to employ a _cluster_
autoscaler that adds and removes Kubernetes _nodes_ according to load. Coder can coexist with such cluster autoscalers,
but we recommend you take steps to prevent the autoscaler from evicting coderd pods, as an eviction will cause the same
interruptions as described above. For example, if you are using the [Kubernetes cluster
autoscaler](https://kubernetes.io/docs/reference/labels-annotations-taints/#cluster-autoscaler-kubernetes-io-safe-to-evict),
you may wish to set `cluster-autoscaler.kubernetes.io/safe-to-evict: "false"` as an annotation on the coderd deployment.

## Troubleshooting

If a load test fails or if you are experiencing performance issues during day-to-day use, you can leverage Coder's [prometheus metrics](./prometheus.md) to identify bottlenecks during scale tests. Additionally, you can use your existing cloud monitoring stack to measure load, view server logs, etc.