Clustering#

LXD can be run in clustering mode, where any number of LXD servers share the same distributed database and can be managed uniformly using the lxc client or the REST API.

Note that this feature was introduced as part of the API extension “clustering”.

Forming a cluster#

First you need to choose a bootstrap LXD node. It can be an existing LXD server or a brand new one. Then you need to initialize the bootstrap node and join further nodes to the cluster. This can be done interactively or with a preseed file.

Note that all further nodes joining the cluster must have identical configuration to the bootstrap node, in terms of storage pools and networks. The only configuration that can be node-specific are the source and size keys for storage pools and the bridge.external_interfaces key for networks.

It is strongly recommended that the number of nodes in the cluster be at least three, so the cluster can survive the loss of at least one node and still be able to establish quorum for its distributed state (which is kept in a SQLite database replicated using the Raft algorithm). If the number of nodes is less than three, then only one node in the cluster will store the SQLite database. When the third node joins the cluster, both the second and third nodes will receive a replica of the database.

Interactively#

Run lxd init and answer yes to the very first question (“Would you like to use LXD clustering?”). Then choose a name for identifying the node, and an IP or DNS address that other nodes can use to connect to it, and answer no to the question about whether you’re joining an existing cluster. Finally, optionally create a storage pool and a network bridge. At this point your first cluster node should be up and available on your network.

You can now join further nodes to the cluster. Note however that these nodes should be brand new LXD servers, or alternatively you should clear their contents before joining, since any existing data on them will be lost.

There are two ways to add a member to an existing cluster; using the trust password or using a join token. A join token for a new member is generated in advance on the existing cluster using the command:

lxc cluster add <new member name>

This will return a single-use join token which can then be used in the join token question stage of lxd init. The join token contains the addresses of the existing online members, as well as a single-use secret and the fingerprint of the cluster certificate. This reduces the amount of questions you have to answer during lxd init as the join token can be used to answer these questions automatically.

Alternatively you can use the trust password instead of using a join token.

To add an additional node, run lxd init and answer yes to the question about whether to use clustering. Choose a node name that is different from the one chosen for the bootstrap node or any other nodes you have joined so far. Then pick an IP or DNS address for the node and answer yes to the question about whether you’re joining an existing cluster.

If you have a join token then answer yes to the question that asks if you have a join token and then copy it in when it asks for it.

If you do not have a join token, but have a trust password instead then, then answer no to the question that asks if you have a join token. Then pick an address of an existing node in the cluster and check the fingerprint that gets printed matches the cluster certificate of the existing members.

Per-server configuration#

As mentioned previously, LXD cluster members are generally assumed to be identical systems.

However to accommodate things like slightly different disk ordering or network interface naming, LXD records some settings as being server-specific. When such settings are present in a cluster, any new server being added will have to provide a value for it.

This is most often done through the interactive lxd init which will ask the user for the value for a number of configuration keys related to storage or networks.

Those typically cover:

  • Source device for a storage pool (leaving empty would create a loop)

  • Name for a ZFS zpool (defaults to the name of the LXD pool)

  • External interfaces for a bridged network (empty would add none)

  • Name of the parent network device for managed physical or macvlan networks (must be set)

It’s possible to lookup the questions ahead of time (useful for scripting) by querying the /1.0/cluster API endpoint. This can be done through lxc query /1.0/cluster or through other API clients.

Preseed#

Create a preseed file for the bootstrap node with the configuration you want, for example:

config:
  core.trust_password: sekret
  core.https_address: 10.55.60.171:8443
  images.auto_update_interval: 15
storage_pools:
- name: default
  driver: dir
networks:
- name: lxdbr0
  type: bridge
  config:
    ipv4.address: 192.168.100.14/24
    ipv6.address: none
profiles:
- name: default
  devices:
    root:
      path: /
      pool: default
      type: disk
    eth0:
      name: eth0
      nictype: bridged
      parent: lxdbr0
      type: nic
cluster:
  server_name: node1
  enabled: true

Then run cat <preseed-file> | lxd init --preseed and your first node should be bootstrapped.

Now create a bootstrap file for another node. You only need to fill in the cluster section with data and configuration values that are specific to the joining node.

Be sure to include the address and certificate of the target bootstrap node. To create a YAML-compatible entry for the cluster_certificate key you can use a command like sed ':a;N;$!ba;s/\n/\n\n/g' /var/lib/lxd/cluster.crt (or sed ':a;N;$!ba;s/\n/\n\n/g' /var/snap/lxd/common/lxd/cluster.crt for snap users), which you have to run on the bootstrap node. cluster_certificate_path key (which should contain valid path to cluster certificate) can be used instead of cluster_certificate key.

For example:

cluster:
  enabled: true
  server_name: node2
  server_address: 10.55.60.155:8443
  cluster_address: 10.55.60.171:8443
  cluster_certificate: "-----BEGIN CERTIFICATE-----

opyQ1VRpAg2sV2C4W8irbNqeUsTeZZxhLqp4vNOXXBBrSqUCdPu1JXADV0kavg1l

2sXYoMobyV3K+RaJgsr1OiHjacGiGCQT3YyNGGY/n5zgT/8xI0Dquvja0bNkaf6f

...

-----END CERTIFICATE-----
"
  cluster_password: sekret
  member_config:
  - entity: storage-pool
    name: default
    key: source
    value: ""

When joining a cluster using a cluster join token, the following fields can be omitted:

  • server_name

  • cluster_address

  • cluster_certificate

  • cluster_password

And instead the full token be passed through the cluster_token field.

Managing a cluster#

Once your cluster is formed you can see a list of its nodes and their status by running lxc cluster list. More detailed information about an individual node is available with lxc cluster show <node name>.

Cluster member configuration#

Each cluster member has its own key/value configuration with the following supported namespaces:

  • user (free form key/value for user metadata)

  • scheduler (options related to how the member is automatically targeted by the cluster)

The currently supported keys are:

Key

Type

Default

Description

scheduler.instance

string

all

If all then the member will be auto-targeted for instance creation if it has the least number of instances. If manual then instances will only target the member if --target is given. If group then instances will only target members in the group provided using --target=@<group>

user.*

string

-

Free form user key/value storage (can be used in search)

Cluster member roles#

The following roles can be assigned to LXD cluster members. Automatic roles are assigned by LXD itself and cannot be modified by the user.

Role

Automatic

Description

database

yes

Voting member of the distributed database

database-leader

yes

Current leader of the distributed database

database-standby

yes

Stand-by (non-voting) member of the distributed database

event-hub

no

Exchange point (hub) for the internal LXD events (requires at least two)

ovn-chassis

no

Uplink gateway candidate for OVN networks

Voting and stand-by members#

The cluster uses a distributed database to store its state. All nodes in the cluster need to access such distributed database in order to serve user requests.

If the cluster has many nodes, only some of them will be picked to replicate database data. Each node that is picked can replicate data either as “voter” or as “stand-by”. The database (and hence the cluster) will remain available as long as a majority of voters is online. A stand-by node will automatically be promoted to voter when another voter is shutdown gracefully or when its detected to be offline.

The default number of voting nodes is 3 and the default number of stand-by nodes is 2. This means that your cluster will remain operation as long as you switch off at most one voting node at a time.

You can change the desired number of voting and stand-by nodes with:

lxc config set cluster.max_voters <n>

and

lxc config set cluster.max_standby <n>

with the constraint that the maximum number of voters must be odd and must be least 3, while the maximum number of stand-by nodes must be between 0 and 5.

Deleting nodes#

To cleanly delete a node from the cluster use lxc cluster remove <node name>.

Offline nodes and fault tolerance#

At each time there will be an elected cluster leader that will monitor the health of the other nodes. If a node is down for more than 20 seconds, its status will be marked as OFFLINE and no operation will be possible on it, as well as operations that require a state change across all nodes.

If the node that goes offline is the leader itself, the other nodes will elect a new leader.

As soon as the offline node comes back online, operations will be available again.

If you can’t or don’t want to bring the node back online, you can delete it from the cluster using lxc cluster remove --force <node name>.

You can tweak the amount of seconds after which a non-responding node will be considered offline by running:

lxc config set cluster.offline_threshold <n seconds>

The minimum value is 10 seconds.

Upgrading nodes#

To upgrade a cluster you need to upgrade all of its nodes, making sure that they all upgrade to the same version of LXD.

To upgrade a single node, simply upgrade the lxd/lxc binaries on the host (via snap or other packaging systems) and restart the LXD daemon.

If the new version of the daemon has database schema or API changes, the restarted node might transition into a Blocked state. That happens if there are still nodes in the cluster that have not been upgraded and that are running an older version. When a node is in the Blocked state it will not serve any LXD API requests (in particular, lxc commands on that node will not work, although any running instance will continue to run).

You can see if some nodes are blocked by running lxc cluster list on a node which is not blocked.

As you proceed upgrading the rest of the nodes, they will all transition to the Blocked state, until you upgrade the very last one. At that point the blocked nodes will notice that there is no out-of-date node left and will become operational again.

Evacuating and restoring cluster members#

Whether it’s for routine maintenance like applying system updates requiring a reboot or to perform hardware changes, you may sometimes want to empty a given server of all its instances.

This can be done using lxc cluster evacuate <NAME> which will migrate all instances on that server, moving them to other cluster members. The evacuated cluster member will be transitioned to an “evacuated” state which will prevent the creation of any instances on it.

Once maintenance is complete, lxc cluster restore <NAME> will move the server back into a normal running state and will move its instances back from the servers that were temporarily holding them.

The behavior for a given instance can be configured through the cluster.evacuate instance configuration key. Instances will be shutdown cleanly, respecting the boot.host_shutdown_timeout configuration key.

Failure domains#

Failure domains can be used to indicate which nodes should be given preference when trying to assign roles to a cluster member that has been shutdown or has crashed. For example, if a cluster member that currently has the database role gets shutdown, LXD will try to assign its database role to another cluster member in the same failure domain, if one is available.

To change the failure domain of a cluster member you can use the lxc cluster edit <member> command line tool, or the PUT /1.0/cluster/<member> REST API.

Recover from quorum loss#

Every LXD cluster has up to 3 members that serve as database nodes. If you permanently lose a majority of the cluster members that are serving as database nodes (for example you have a 3-member cluster and you lose 2 members), the cluster will become unavailable. However, if at least one database node has survived, you will be able to recover the cluster.

In order to check which cluster members are configured as database nodes, log on any survived member of your cluster and run the command:

lxd cluster list-database

This will work even if the LXD daemon is not running.

Among the listed members, pick the one that has survived and log into it (if it differs from the one you have run the command on).

Now make sure the LXD daemon is not running and then issue the command:

lxd cluster recover-from-quorum-loss

At this point you can restart the LXD daemon and the database should be back online.

Note that no information has been deleted from the database, in particular all information about the cluster members that you have lost is still there, including the metadata about their instances. This can help you with further recovery steps in case you need to re-create the lost instances.

In order to permanently delete the cluster members that you have lost, you can run the command:

lxc cluster remove <name> --force

Note that this time you have to use the regular lxc command line tool, not lxd.

Recover cluster members with changed addresses#

If some members of your cluster are no longer reachable, or if the cluster itself is unreachable due to a change in IP address or listening port number, the cluster can be reconfigured.

On each member of the cluster, with LXD not running, run the following command:

lxd cluster edit

Note that all commands in this section will use lxd instead of lxc.

This will present a YAML representation of this node’s last recorded information about the rest of the cluster:

# Latest dqlite segment ID: 1234

members:
  - id: 1             # Internal ID of the node (Read-only)
    name: node1       # Name of the cluster member (Read-only)
    address: 10.0.0.10:8443 # Last known address of the node (Writeable)
    role: voter             # Last known role of the node (Writeable)
  - id: 2
   name: node2
    address: 10.0.0.11:8443
    role: stand-by
  - id: 3
   name: node3
    address: 10.0.0.12:8443
    role: spare

Members may not be removed from this configuration, and a spare node cannot become a voter, as it may lack a global database. Importantly, keep in mind that at least 2 nodes must remain voters (except in the case of a 2-member cluster, where 1 voter suffices), or there will be no quorum.

Once the necessary changes have been made, repeat the process on each member of the cluster. Upon reloading LXD on each member, the cluster in its entirety should be back online with all nodes reporting in.

Note that no information has been deleted from the database, all information about the cluster members and their instances is still there.

Instances#

You can launch an instance on any node in the cluster from any node in the cluster. For example, from node1:

lxc launch --target node2 ubuntu:22.04 c1

will launch an Ubuntu 22.04 container on node2.

When you launch an instance without defining a target, the instance will be launched on the server which has the lowest number of instances. If all the servers have the same amount of instances, it will choose one at random.

You can list all instances in the cluster with:

lxc list

The NODE column will indicate on which node they are running.

After an instance is launched, you can operate it from any node. For example, from node1:

lxc exec c1 ls /
lxc stop c1
lxc delete c1
lxc pull file c1/etc/hosts .

Manually altering Raft membership#

There might be situations in which you need to manually alter the Raft membership configuration of the cluster because some unexpected behavior occurred.

For example if you have a cluster member that was removed uncleanly it might not show up in lxc cluster list but still be part of the Raft configuration (you can see that with lxd sql local "SELECT * FROM raft_nodes").

In that case you can run:

lxd cluster remove-raft-node <address>

to remove the leftover node.

Images#

By default, LXD will replicate images on as many cluster members as you have database members. This typically means up to 3 copies within the cluster.

That number can be increased to improve fault tolerance and likelihood of the image being locally available.

The special value of “-1” may be used to have the image copied on all nodes.

You can disable the image replication in the cluster by setting the count down to 1:

lxc config set cluster.images_minimal_replica 1

Storage pools#

As mentioned above, all nodes must have identical storage pools. The only difference between pools on different nodes might be their source, size or zfs.pool_name configuration keys.

To create a new storage pool, you first have to define it across all nodes, for example:

lxc storage create --target node1 data zfs source=/dev/vdb1
lxc storage create --target node2 data zfs source=/dev/vdc1

Note that when defining a new storage pool on a node the only valid configuration keys you can pass are the node-specific ones mentioned above.

At this point the pool hasn’t been actually created yet, but just defined (it’s state is marked as Pending if you run lxc storage list).

Now run:

lxc storage create data zfs

and the storage will be instantiated on all nodes. If you didn’t define it on a particular node, or a node is down, an error will be returned.

You can pass to this final storage create command any configuration key which is not node-specific (see above).

Storage volumes#

Each volume lives on a specific node. The lxc storage volume list includes a NODE column to indicate on which node a certain volume resides.

Different volumes can have the same name as long as they live on different nodes (for example image volumes). You can manage storage volumes in the same way you do in non-clustered deployments, except that you’ll have to pass a --target <node name> parameter to volume commands if more than one node has a volume with the given name.

For example:

# Create a volume on the node this client is pointing at
lxc storage volume create default web

# Create a volume with the same node on another node
lxc storage volume create default web --target node2

# Show the two volumes defined
lxc storage volume show default web --target node1
lxc storage volume show default web --target node2

Networks#

As mentioned above, all nodes must have identical networks defined.

The only difference between networks on different nodes might be their optional configuration keys. When defining a new network on a specific clustered node the only valid optional configuration keys you can pass are bridge.external_interfaces and parent, as these can be different on each node (see documentation about network configuration for a definition of each).

To create a new network, you first have to define it across all nodes, for example:

lxc network create --target node1 my-network
lxc network create --target node2 my-network

At this point the network hasn’t been actually created yet, but just defined (it’s state is marked as Pending if you run lxc network list).

Now run:

lxc network create my-network

The network will be instantiated on all nodes. If you didn’t define it on a particular node, or a node is down, an error will be returned.

You can pass to this final network create command any configuration key which is not node-specific (see above).

Separate REST API and clustering networks#

You can configure different networks for the REST API endpoint of your clients and for internal traffic between the nodes of your cluster (for example in order to use a virtual address for your REST API, with DNS round robin).

To do that, you need to bootstrap the first node of the cluster using the cluster.https_address configuration key. For example, when using preseed:

config:
  core.trust_password: sekret
  core.https_address: my.lxd.cluster:8443
  cluster.https_address: 10.55.60.171:8443
...

(the rest of the preseed YAML is the same as above).

To join a new node, first set its REST API address, for instance using the lxc client:

lxc config set core.https_address my.lxd.cluster:8443

and then use the PUT /1.0/cluster API endpoint as usual, specifying the address of the joining node with the server_address field. If you use preseed, the YAML payload would be exactly like the one above.

Updating the cluster certificate#

In a LXD cluster, all servers respond with the same shared certificate. This is usually a standard self-signed certificate with an expiry set to 10 years.

If you wish to replace it with something else, for example a valid certificate obtained through Let’s Encrypt, lxc cluster update-certificate can be used to replace the certificate on all servers in your cluster.

Cluster groups#

In a LXD cluster, members can be added to cluster groups. By default, all members belong to the default group.

Cluster members can be assigned to groups using the lxc cluster group assign command:

lxc cluster group create gpu
lxc cluster group assign cluster:node1 gpu

With cluster groups, it’s possible to target specific groups instead of individual members. This is done by using the @ prefix when using --target.

An example:

lxc launch ubuntu:22.04 cluster:ubuntu --target=@gpu

This will cause the instance to be created on a cluster member belonging to gpu group if scheduler.instance is set to either all (default) or group.