---
title: Replace a server
description: "Move a member of an HA service or a DCS cluster to a different server: release the slot, fill it with a free server, and rebuild the member."
sidebar:
  order: 12
search:
  keywords: ["machine", "hardware", "swap a server", "replace a node"]
related:
  - /operations/resize-in-place
  - /database-services/external-dcs
---

## What happens

An HA service has one **member slot** for each PostgreSQL member (`pgnode01`,
`pgnode02`, ...). A DCS cluster has one slot for each etcd member (`etcd01`,
...). **Replace server** (admin) has three steps:

1. **Release.** The server leaves the slot, and the console deletes its
   database data.
2. **Fill.** A free server, or a server that you enroll now, takes the slot
   and its member name.
3. **Rebuild.** The console copies the data from the leader and starts the new
   member as a replica.

The other members continue to serve and do not restart. The rebuild copies
the full database from the leader: schedule it for a quiet period. During a
replacement, fewer members can take over in a failover.

Replace server works on Standard HA, Advanced HA and
[DCS clusters](#dcs-member). For a single-node service,
[restore a backup](/docs/v2026.11.5-beta/backups/restore) to a new service.

## One server at a time

On Standard HA, etcd runs on the three PostgreSQL nodes and needs 2 of 3
members. If two members are lost, the primary demotes, and writes stop. The
console enforces these rules:

- A release or a fill starts only when each other member is healthy.
- While a slot is empty or waits for a rebuild, **Replace server** is disabled
  on the other members (*another member is being replaced*). Fill the empty
  slot first.
- While a slot is empty or waits for a rebuild, or a member is down, the
  console refuses **Deploy**, **Rolling restart**, **Change VIP** and applying
  the license limits. The Overview shows the **DCS quorum at risk** banner.
- If the refusal says *console cannot reach the mesh*, repair the network of
  the console first.
- If two members of a Standard HA service are lost, the refusal says *both
  out, so etcd has no quorum*. Follow
  [Site outage and recovery](/docs/v2026.11.5-beta/operations/site-outage#etcd-has-no-quorum).
- On Advanced HA, if one slot is empty and another member waits for a
  rebuild, rebuild that member first (*rebuild it first, then fill*).

## Before you start

- No other operation runs on the service. A fill also starts on a **Failed**
  service that has a running leader. After the rebuild, the Overview shows
  **Redeploy recommended** until a deploy completes.
- The topology shows one primary, and each other surviving node is a healthy
  replica. If not, do not release or fill the slot. Contact
  [support](/docs/v2026.11.5-beta/logs-and-support).
- The new server meets the
  [server requirements](/docs/v2026.11.5-beta/servers#server-requirements), is not smaller than
  the old server, and is online. It has a new OS install and an empty data
  disk. Do not use a copy of the system disk of another server: the copy has
  the mesh identity of that server.
- On a service with a VIP, the new server has an IPv4 address in the subnet
  of the VIP.
- You have a recent [backup](/docs/v2026.11.5-beta/backups).

## Replace a healthy server

Use this procedure to move a member to a different server, for example new
hardware.

1. **Open Replace server**

    On the service **Overview**, open the menu of the member on the topology
    and select **Replace server**. The member can be a replica or the
    primary.

    ![Replace server panel with the Release, Fill and Rebuild steps](/docs/screenshots/v2026.11.5-beta/replace-a-server/replace-server-panel.light.webp)

2. **Release the server**

    Select **Release** *server-name*, read the dialog, and select **Delete the
    data and release**. If the member is the primary, the console first
    switches over, and writes fail for a few seconds. **The console deletes
    the PostgreSQL data directory, and the etcd data directory on Standard
    HA. You cannot undo this.**

    Wait until the topology shows **pgnode02: no server** and the **Servers**
    page shows the server as **Free**.

3. **Fill the slot**

    Select **Use a free server** and a server, or **Enroll a new server** (see
    [Fill an empty slot](#fill-an-empty-slot)). Select **Fill with the
    selected server**. You can select the server that you released.

4. **Rebuild**

    The fill starts the rebuild. The panel shows the progress. At the end,
    the topology shows the member as a healthy replica. If the rebuild fails,
    see [Rebuild after a reinstall](#rebuild-after-a-reinstall).

If the release fails, the panel shows **The release failed**, and the server
stays in its slot. Read the operation log, correct the cause, and release
again. The release refuses if the server is also a backup repository or a
member of another service or DCS cluster. Remove that use first.

### LiveSync sessions

- A session on this service moves to the primary and shows **Deploying** for
  a short time. The rebuild installs LiveSync on the new member.
- A session created on an early access release (`v0.X.Y-eap.N`) can block the
  release, and the refusal names it. Stop and start the session, then release
  again. If its server is lost, delete the session, replace the server, and
  create a new session. The new session does not resume the old replication.
  See
  [Sessions created on an early access release](/docs/v2026.11.5-beta/livesync/operations#sessions-created-on-an-early-access-release).
- If the LiveSync setup fails, the member continues to wait for a rebuild.
  Correct the cause and retry the rebuild. Then stop and start a **Running**
  session, or start a **Failed** session.

## Failed server

Use this procedure when a server is down: a failed disk, a lost VM, or a
server that you cannot reach. If the lost member was the primary, wait until
the topology shows a new primary.

:::warning
When the console can reach the released server again, it deletes its
PostgreSQL and etcd data. If you need data that is only on the old server,
keep the server isolated and contact support before you bring it back.
:::

1. **Open Replace server**

    On the service **Overview**, the member shows as unreachable. Open its
    menu on the topology and select **Replace server**.

2. **Release the server**

    The dialog says *The server does not answer.* Select **Release the
    server**. The **Servers** page shows the server as **Unreachable** and
    **Needs cleanup**.

3. **Check the leader and fill the slot**

    Make sure that the topology shows a primary: the rebuild copies from it.
    If there is no primary, do not fill the slot. Contact
    [support](/docs/v2026.11.5-beta/logs-and-support). Otherwise, fill the slot as in
    [Fill an empty slot](#fill-an-empty-slot).

4. **Rebuild**

    The fill starts the rebuild. Wait until the topology shows the member as
    a healthy replica. If the rebuild fails, see
    [Rebuild after a reinstall](#rebuild-after-a-reinstall).

Keep the old server powered off: until the cleanup runs, it can start its old
services. When it is online again, the console deletes its data, and the
server shows **Free**. If the cleanup fails, the server stays **Needs
cleanup** with the reason. The console tries again later, or you select
**Retry cleanup** on the server page. If you delete an unreachable server on
the **Servers** page, its services do not stop, and the data on its disk
stays.

## Fill an empty slot

![The card of an empty slot in the topology, with Fill with a server](/docs/screenshots/v2026.11.5-beta/replace-a-server/empty-slot-card.light.webp)

Select **Fill with a server** on the card, or continue in the **Replace
server** panel.

- **Use a free server.** The panel lists the free servers with their LAN
  address and size. You cannot select an unreachable server.
- **Enroll a new server.** Register the server in the panel and run the
  enroll command on it. When the server is online, select it. If enrollment
  fails, see [Enroll a server](/docs/v2026.11.5-beta/servers#enroll-a-server).

The rebuild replaces all PostgreSQL and etcd data on the server.

## Rebuild after a reinstall

To reinstall the OS or replace the disk of a server and keep it in its slot,
do not release it:

1. **Reinstall and re-enroll**

    Reinstall the server. On the **Servers** page, open its menu, select
    **Re-enroll command**, and run the command on the server. The slot then
    waits for a rebuild.

2. **Rebuild**

    On the topology, open the menu of the member, select **Replace server**,
    then select **Rebuild**. Wait until the topology shows the member as a
    healthy replica.

If a rebuild fails, the panel shows **The rebuild failed** and the reason.
The other members continue to serve. Correct the cause and select **Retry
the rebuild**. The retry continues from the failed step.

## DCS member

On Advanced HA, etcd runs on the three servers of a
[DCS cluster](/docs/v2026.11.5-beta/database-services/external-dcs). The DCS cluster needs 2 of 3
members. The rules of [One server at a time](#one-server-at-a-time) apply to
each service on the cluster. If two DCS members are lost, the refusal says
*another member is down (etcd quorum lost)*. Follow
[Site outage and recovery](/docs/v2026.11.5-beta/operations/site-outage#etcd-has-no-quorum).

Before you start, make sure that the detail page of the DCS cluster shows the
two other members as healthy. Each service on the cluster must meet
[Before you start](#before-you-start).

1. **Open Replace server**

    Open **Infrastructure**, then **DCS clusters**, and open the cluster. On
    the topology, open the menu of the member and select **Replace server**.

2. **Release the server**

    Release the server as for a service member. If the server answers, the
    console deletes its etcd data. If it does not answer, read the warning in
    [Failed server](#failed-server). The slot shows **etcd03: no server**.

3. **Fill the slot**

    Fill the slot as in [Fill an empty slot](#fill-an-empty-slot).

4. **Rebuild**

    The fill starts the rebuild. With the default settings, it ends about
    five minutes after the new member joins. Wait until the cluster shows
    **Healthy** before you replace another member.

If the rebuild fails, correct the cause and select **Retry the rebuild**.

## After the replacement

- **Connection endpoints.** With a VIP, clients do not change. Without a VIP,
  the direct endpoint of the member changes if the LAN address changes.
  Update those clients. The connection info of the service shows the new
  address.
- **A larger server.** The new member has the PostgreSQL settings of the
  leader. To tune them for the new size, open **Database service admin** and
  select **Retune** under **Operations**. Retune restarts PostgreSQL, with one
  switchover.
- **The released server.** When it is **Free**, use it for another slot or a
  new service, or delete it. You cannot delete it until the slot that it left
  completes its rebuild.
- **The license.** The release returns the license seat of the old server,
  and the fill takes a seat for the new server. This also works after the
  license expires. DCS servers do not use license seats.
- **The backup server.** The public key of the old server stays in
  `~postgres/.ssh/authorized_keys` on the backup server. You can delete that
  line.
