# Drain a site

> Move traffic away from a server before maintenance, without a hard withdraw.

Draining (RFC 8326) takes a server out of an anycast set without a hard withdraw: our routes
go out again with the GRACEFUL_SHUTDOWN community `65535:0`, so routers move traffic to the
other sites while this one still answers.

```sh
ansage drain --wait 30 && systemctl stop ansage
```

This drains for 30 s, then withdraws. `drain_time` does the same on every stop (a second
SIGTERM ends the wait).

## What changes on the wire

| Session | ansage sends                                                     |
| ------- | ---------------------------------------------------------------- |
| iBGP    | GRACEFUL_SHUTDOWN, LOCAL_PREF 0                                  |
| eBGP    | GRACEFUL_SHUTDOWN, plus `drain_prepend` extra prepends of our AS |

## Router side

**FRR:**

FRR lowers routes with `65535:0` to LOCAL_PREF 0 by itself. Nothing to configure.

**BIRD:**

BIRD needs a filter, or set `drain_prepend` on the ansage side:

```txt
if (65535, 0) ~ bgp_community then bgp_local_pref = 0;
```

## Settings

```toml
drain_time = 30        # on SIGTERM/SIGINT, drain this many seconds before closing (default 0)

[[neighbor]]
address = "10.0.0.1"
asn = 64501
drain_prepend = 2      # eBGP, for neighbors that ignore GRACEFUL_SHUTDOWN
```

## Is it drained?

```txt
$ ansage show summary
...
Announcing: 192.0.2.0/24 2001:db8:1::/48
Drained for 42.0s: our routes carry GRACEFUL_SHUTDOWN (ansage undrain ends it)
```

The metric `ansage_drained` is 1 while drained; `/api/summary` has the drain start in `drained`.

> **Lifetime of a drain:** A drain lasts until `ansage undrain`, survives a reload and ends with a restart or upgrade.