# Health checks

How each service of a load balancer probes its targets, the settings and their ranges by type, and how a target leaves and rejoins rotation.

A health check belongs to a service, not to the load balancer: every service probes every target
on its own. A target that fails a service's check stops receiving that service's traffic and gets
it again once it passes.

## Settings

| Setting               | Values                                                                                                                             | Default         |
| --------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | --------------- |
| `healthCheckProtocol` | `http`: a GET of the path, 2xx or 3xx is healthy. `tcp`: a connection is enough.                                                   | `http`          |
| `healthCheckPath`     | the path to `GET`; when changed later, it must start with `/` and have at most 256 characters (adding a service does not check it) | `/`             |
| `healthCheckInterval` | seconds between checks: 3–60 on `lb11`/`lb21`/`lb31`, 1–60 on `application` and `dedicated`                                        | 15              |
| `healthCheckTimeout`  | seconds a check may take, 1–60 and at most the interval                                                                            | 10              |
| `healthCheckRetries`  | failed checks in a row before a target stops taking traffic, 0–5                                                                   | 3               |
| `healthCheckPort`     | `dedicated` only: the port to probe instead of the target port                                                                     | the target port |

The check is made against the service's destination port on each target, unless a `dedicated`
service sets `healthCheckPort`. The console's **Add service** form sets only the path; the other
settings are set with the CLI, API or Terraform, when the service is added or later. See
[Change a health check or sticky sessions](change-a-health-check.md).

## On `application` and `dedicated`

The fleet's machines run the check. A target leaves rotation after `healthCheckRetries` failed
checks in a row (at least one) and comes back after two passing ones. Each machine of the fleet
checks every target itself. The target's `status` in the API and the console is the result the
fleet's active machine reports: `healthy` or `unhealthy`. Before its first report, a new target
shows `unknown` on `application` and `healthy` on `dedicated`. Either way, a new target takes traffic
as soon as the fleet carries it, until the check takes it out. The **Live** tab shows each target's health as each
machine sees it.

On `http` and `https` services, a `GET`, `HEAD` or `OPTIONS` request without a body that cannot
connect to its target is tried once on another target.

## On `lb11`, `lb21` and `lb31`

The provider runs the check with the settings of the service. A target's `status` in the API and
the console is not the check's result: it stays `healthy`.

## Close connections on a failed check (`dedicated`, tcp)

With `closeOnUnhealthy: true` on a `dedicated` `tcp` service, the open connections to a target
are closed when its health check answers no (an HTTP status outside 2xx and 3xx, `healthCheckRetries`
times in a row), so clients reconnect to another target. A check that only times out stops new
connections to the target and closes nothing. This suits a database primary endpoint, with an
HTTP check on a port that answers 2xx only on the primary.

## Related

- [Change a health check or sticky sessions](change-a-health-check.md)
- [Add or remove a service](add-or-remove-a-service.md)
- [Troubleshooting](troubleshooting.md)
