# Rolling deploys no longer interrupt running agents

> Servers drain connections and workers hand running turns to another worker on shutdown, so a rolling restart no longer cuts streams or stalls turns.

Source: <https://everruns.com/docs/changelog/2026-10-29-graceful-rolling-deploys/>

Date: 2026-10-29
Products: Platform, Cloud
Category: Improvement
Available: In 0.50.1

Restarting a server now drains it. It keeps serving for a few seconds while the load balancer moves traffic away, then ends event streams with a `disconnecting` event (reason `shutdown`) so clients reconnect to another replica, and lets in-flight requests finish. Tune it with [`SERVER_SHUTDOWN_DELAY_SECS` and `SERVER_SHUTDOWN_GRACE_SECS`](https://everruns.com/docs/sre/environment-variables/#server-shutdown).

```text
event: disconnecting
data: {"reason":"shutdown","retry_ms":1000}
```

Stopping a worker now hands each running turn to another worker at its next step, instead of holding it until the worker is killed and its lease expires. The turn continues within about a second, and finished tool calls do not run again.

Requests that hit a database timeout while authenticating now get `503 Service Unavailable` with `Retry-After`, not `401 Unauthorized`.

The [Helm chart](https://everruns.com/docs/getting-started/kubernetes/#upgrade) sets these defaults. Elsewhere, keep the orchestrator’s kill timeout above these grace periods:

```yaml
# Server: 5 s linger + 20 s drain + 5 s worker link
terminationGracePeriodSeconds: 45
```