Important

This feature is only available in Canton Enterprise

High Availability Usage

Overview

Canton nodes can be deployed in a highly available manner to ensure that domains and participants will continue operating despite isolated machine failures. See High Availability for a detailed description of the architecture in each Canton component to support HA.

Domain Manager

As explained in Domain Architecture and Integrations, a domain internally comprises a sequencer, a mediator and a topology manager. When running a simple domain node (configured with canton.domains, as shown in most of the examples), this node will be running a topology manager, a sequencer and a mediator all internally.

It is possible however to run sequencer(s) and mediator(s) as standalone nodes, as will be explained in the next topics. But to complete the domain setup, it is also necessary to run a domain manager node (configured with canton.domain-managers), which takes care of the bootstrapping of the distributed domain setup and runs the topology manager.

The domain bootstrapping process is explained in Domain bootstrapping.

The domain manager node itself is currently not HA but it is not on the critical path for transaction processing, but for onboarding new parties/participants.

HA Setup on Oracle

The HA approach that is used by the participant, mediator, and sequencer nodes requires additional permissions being granted on Oracle to the database user.

All replicas of a node must be configured with the same DB user name. The DB user must have the following permissions granted:

GRANT EXECUTE ON SYS.DBMS_LOCK TO $username
GRANT SELECT ON V_$LOCK TO $username
GRANT SELECT ON V_$MYSTAT TO $username

In the above commands the $username must be replaced with the configured DB user name. These permissions allow the DB user to request application-level locks on Oracle, as well as to query the state of locks and its own session information.

Mediator

The mediator service uses a hot-standby mechanism, with an arbitrary number of replicas.

Running a Stand-Alone Mediator Node

A domain may be statically configured with a single embedded mediator node or it may be configured to work with external mediators. Once the domain has been initialized further mediators can be added at runtime.

By default a domain node will run an embedded mediator node itself. This is useful in simple deployments where all domain functionality can be co-located on a single host. In a distributed setup where domain services are operated over many machines you can instead configure a domain manager node and bootstrap the domain with mediator(s) running externally.

Mediator nodes can be defined the same manner as Canton participants and domains.

mediators {
  mediator1 {
    admin-api.port = 5017
  }

When the domain node starts it will automatically provide the embedded mediator information about the domain. External mediators have to be initialized using runtime administration in order to complete the domains initialization.

HA Configuration

HA mediator support is only available in the Enterprise version of Canton and only PostgreSQL and Oracle based storage are supported for HA.

Mediator node replicas are configured in the Canton configuration file as individual stand-alone mediator nodes with two required changes for each mediator node replica:

  • Using the same storage configuration to ensure access to the shared database.
  • Set replication.enabled = true for each mediator node replica.

Only the active mediator node replica has to be initialized through the domain bootstrap commands. The passive replicas observe the initialization via the shared database.

Further replicas can be started at runtime without any additional setup. They remain passive until the current active mediator node replica fails.

Sequencer

The database based sequencer can be horizontally scaled and placed behind a load-balancer to provide high availability and performance improvements.

Deploy multiple sequencer nodes for the Domain with the following configuration:

  • All sequencer nodes share the same database so ensure that the storage configuration for each sequencer matches.
  • All sequencer nodes must be configured with high-availability.enabled = true.
canton {
  sequencers {
    sequencer1 {
      sequencer {
        type = database
        high-availability.enabled = true
      }

The Domain node only supports embedded sequencers, so a distributed setup using a domain manager node must then be configured to use these Sequencer nodes by pointing it at these external services.

Once configured the domain must be bootstrapped with the new external sequencer using the bootstrap_domain operational process. These sequencers share a database so just use a single instance for bootstrapping and the replicas will come online once the shared database has sufficient state for starting.

As these nodes are likely running in separate processes you could run this command entirely externally using a remote administration configuration.

canton {
  remote-domains {
    da {
      # these details are provided to other nodes to use for how they should connect to the embedded sequencer
      public-api {
        address = da-domain.local
        port = 1234
      }
      admin-api {
        address = da-domain.local
        port = 1235
      }
    }
  }

  remote-sequencers {
    sequencer1 {
      # these details are provided to other nodes to use for how they should connect to the sequencer
      public-api {
        address = sequencer1.local
        port = 1235
      }
      # the server used from running administration commands
      admin-api {
        address = sequencer1.local
        port = 1235
      }
    }
  }
}

There are two methods available for exposing the horizontally scaled sequencer instances to participants.

Total Node Count

The sequencer.high-availability.total-node-count parameter is used to divide up time among the database sequencers. Because each message sequenced must have a unique timestamp, a sequencer node will use timestamps modulo the total-node-count plus own index in order to create timestamps that do not conflict with other sequencer nodes while sequencing the messages in a parallel database insertion process. Canton uses microseconds, which yields a theoretical max throughput of 1 million messages per second per domain. Now, this theoretical throughput is divided equally among all sequencer nodes (total-node-count). Therefore, if you set total-node-count too high, then a sequencer might not be able to operate at the maximum theoretical throughput. We recommend to keep the default value of 10, as all above explanations are only of theoretical nature and we have not yet seen a database / hard-disk that can handle the theoretical throughput. Also note that a message might contain multiple events, such that we are talking about high numbers here.

External load balancer

Using a load balancer is recommended when you have a http2+grpc supporting load balancer available, and can’t/don’t want to expose details of the backend sequencers to clients. An advanced deployment could also support elastically scaling the number of sequencers available and dynamically reconfigure the load balancer for this updated set.

An example HAProxy configuration for exposing GRPC services without TLS looks like:

frontend domain_frontend
  bind 1234 proto h2
  default_backend domain_backend

backend domain_backend
  balance roundrobin
  server sequencer1 sequencer1.local:1234 proto h2
  server sequencer2 sequencer2.local:1234 proto h2
  server sequencer3 sequencer3.local:1234 proto h2

Client-side load balancing

Using client-side load balancing is recommended where a external load-balancing service is unavailable (or lacks http2+grpc support), and the set of sequencers is static and can be configured at the client.

To simply specify multiple sequencers use the domains.connect_ha console command when registering/connecting to the domain:

myparticipant.domains.connect_ha(
  "my_domain_alias",
  "https://sequencer1.example.com",
  "https://sequencer2.example.com",
  "https://sequencer3.example.com"
)

See the documentation on the connect command using a domain connection config for how to add many sequencer urls when combined with other domain connection options. The domain connection configuration can also be changed at runtime to add or replace configured sequencer connections. Note the domain will have to be disconnected and reconnected at the participant for the updated configuration to be used.

Participant

High availability of a participant node is achieved by running multiple participant node replicas that have access to a shared database.

Participant node replicas are configured in the Canton configuration file as individual participants with two required changes for each participant node replica:

  • Using the same storage configuration to ensure access to the shared database. Only PostgreSQL and Oracle based storage is supported for HA. For Oracle it is crucial that the participant replicas use the same username to access the shared database.
  • Set replication.enabled = true for each participant node replica.

Domain Connectivity during Fail-over

During fail-over from one replica to another the new active replica re-connects to all configured domains for which manualConnect = false. This means if the former active replica was manually connected to a domain, this domain connection is not automatically re-established during fail-over, but must be performed manually again.

Manual Trigger of a Fail-over

Fail-over from the active to a passive replica is done automatically when the active replica has a failure, but one can also initiate a graceful fail-over with the following command:

activeParticipantReplica.replication.set_passive()

The command succeeds if there is at least another passive replica that takes over from the current active replica, otherwise the active replica remains active.

Load Balancer Configuration

Many replicated participants can be placed behind an appropriately sophisticated load balancer that will by health checks determine which participant instance is active and direct ledger and admin api requests to that instance appropriately. This makes participant replication and failover transparent from the perspective of the ledger-api application or canton console administering the logical participant, as they will simply be pointed at the load balancer.

Participants should be configured to expose an “IsActive” health status on our health http server using the following monitoring configuration:

canton {
  monitoring {
    health {
      server {
        address = 0.0.0.0
        port = 8000
      }

      check.type = is-active
    }
  }
}

Once running this server will report a http 200 status code on a http/1 GET request to /health if the participant is currently the active replica. Otherwise an error will be returned.

To use a load balancer it must support http/1 health checks for routing requests on a separate http/2 (GRPC) server. This is possible with HAProxy using the following example configuration:

global
    log stdout format raw local0

defaults
    log global
    mode http
    option httplog
    # enabled so long running connections are logged immediately upon connect
    option logasap

# expose the admin-api and ledger-api as separate servers
frontend admin-api
    bind :15001 proto h2
    default_backend admin-api

backend admin-api
    # enable http health checks
    option httpchk
    # required to create a separate connection to query the load balancer.
    # this is particularly important as the health http server does not support h2
    # which would otherwise be the default.
    http-check connect
    # set the health check uri
    http-check send meth GET uri /health

    # list all participant backends
    server participant1 participant1.lan:15001 proto h2 check port 8080
    server participant2 participant2.lan:15001 proto h2 check port 8080
    server participant3 participant3.lan:15001 proto h2 check port 8080

# repeat a similar configuration to the above for the ledger-api
frontend ledger-api
    bind :15000 proto h2
    default_backend ledger-api

backend ledger-api
    option httpchk
    http-check connect
    http-check send meth GET uri /health

    server participant1 participant1.lan:15000 proto h2 check port 8080
    server participant2 participant2.lan:15000 proto h2 check port 8080
    server participant3 participant3.lan:15000 proto h2 check port 8080