Skip to main content

Features

Three things decide what a cluster can do: the deployment model you pick, how you let applications reach it, and the engine you pick. None of them is easy to change afterwards, so read all three before you provision, and see Product Introduction for the engine catalog itself.

note

Engine and feature availability changes as the service develops. Treat this page as the documented state and the Console Portal as the current one.

Deployment models

Before the per-feature detail, the choice that shapes most of it. Every cluster is deployed in one of two models, picked at provisioning time based on what you need from availability, performance, and cost.

Single NodeHigh Availability (HA / Cluster)
NodesOne database instance, no standbyA primary plus one or more standby or replica nodes
PlacementSingle nodeSame or different Availability Zones
FailoverNoneAutomatic. The database is monitored continuously and fails over to a standby when the primary fails
DowntimeExpected during node failure and maintenanceMinimized by failover
Suited toDevelopment and test, low to medium load, and cases where a simpler and cheaper architecture matters moreProduction and business-critical systems

A Single Node cluster gets the full compute allocation, meaning vCPU, RAM, and storage, on that one node. Backups still run on whatever schedule the service is configured with, so the model is not unprotected. What it lacks is a standby, so availability tracks the health of that single node.

High Availability buys three things beyond reduced downtime: better reliability and recoverability, read scaling through replicas when you configure it, and the option to put Database proxy in front for connection routing.

Which model you can choose, and what HA actually delivers, both vary by engine. The tables below carry the specifics.

Network connectivity

Applications never address a node directly. They connect through a database endpoint, and what sits behind that endpoint follows from the deployment model you chose above.

Deployment modelThe endpoint points to
Single NodeThe database instance
High availabilityA DB Proxy representing the cluster

Going through an endpoint decouples application logic from cluster architecture, so a topology change or a failover does not force you to change connection settings.

Private and public access

Two access models are available, and the choice is a security decision as much as a networking one.

Access modelWhat it meansUse it for
Private accessThe cluster is reachable only inside the private network (VPC)Production, and anywhere exposure has to stay minimal
Public accessThe cluster exposes a public endpoint reachable from the Internet, with additional security controlsExternal systems, temporary administrative access, testing and integration

Security Group and Floating IP

These two controls are what make public access safe enough to use, and they are meant to be used together rather than separately.

A Security Group is a set of stateful firewall rules that allow or deny inbound and outbound traffic by IP address or CIDR, port, and protocol. It is how you enforce least privilege, permitting access only from authorized sources such as applications in the same VPC or an approved bastion host. Rule changes take effect immediately and never require a cluster restart.

A Floating IP is a public IP address you can attach to or detach from a database instance on demand, which is what turns on Internet reachability. Pair it with a Security Group that restricts the allowed source ranges and exposes only the database port you actually need, such as 3306, 5432, or 1433.

warning

Enable a Floating IP only while you need it, and prefer private access for production. A Floating IP with a permissive Security Group puts the database on the public Internet.

The procedure for both is in Connect to database with Floating IP.

Features limited to specific engines

FeatureAvailable onWhat to know
Upgrade database versionMariaDBOne path only, MariaDB to MariaDB 10.9. Other engines show no upgrade control. An upgrade cannot be rolled back. For every other engine, moving between major versions is a manual data migration under the version policy.
Database proxyMySQL, PostgreSQL, MariaDBHigh Availability clusters only. The DB Proxy tab stays hidden unless both conditions hold.
Kafka users and ACLsKafkaSASL authentication for producers, consumers, and Kafka Connect.
Point-in-time recoveryPostgreSQL, MySQL, MariaDBProtection starts only after the next full backup completes, not when you enable it.
Differential backup jobs, in Configure backup schedulePostgreSQL, MySQL, MariaDB, SQL ServerRelational engines only. Full backup jobs are not restricted this way.
Push backup data to S3PostgreSQL, MySQL, MariaDB, SQL Server, Redis, MongoDB, TimescaleDB
High Availability, in Create a databaseEvery engine except CassandraCassandra sets node count through Number Of Nodes instead. On ClickHouse, read the replication note below before relying on it.

Provisioning fields that only some engines show

The create-cluster form changes shape with the engine you pick, so a field described in the guide may not appear for you.

FieldShown for
Edition, meaning Enterprise, Standard, or WebSQL Server only
Database NameEvery engine except Redis and Kafka
VHost NameRabbitMQ only
Number Of NodesCassandra only
High AvailabilityEvery engine except Cassandra

Behavior that differs by engine

These are the cases where the same feature behaves differently depending on what you are running. Each one has caused a surprise worth planning around.

EngineWhat differs
PostgreSQL, OpenSearch, ClickHouseA vCPU or RAM change does not take effect until you restart the cluster yourself. Until you do, it keeps running the old configuration and shows a warning icon in the cluster list. Applies to both resource changes and auto scaling.
PostgreSQL HA clusters created before March 13, 2026Enabling PITR restarts the cluster automatically and briefly interrupts service. Enable it off-peak.
KafkaRunning a backup can interrupt service, so schedule Kafka backups for low-traffic periods. Creating or updating a user with SASL/PLAIN restarts the cluster.
ClickHouse with High AvailabilityReplicas are created per shard, but a table replicates only if you create it with a replicated engine from the MergeTree family. A plain MergeTree table stays on the node that received the write. See the ClickHouse - High Availability FAQ.
CassandraNo High Availability toggle. Size the cluster with Number Of Nodes at provisioning time.
warning

Enabling High Availability on ClickHouse is not enough on its own. Check the table engine on every table you rely on, because an unreplicated MergeTree table loses its data if the node holding it fails.

Features that work the same on every engine

Configured identically whichever engine you run, aside from the engine-specific behavior above:

Next steps