Google Cloud: Create a high-availability instance
A high-availability (Pro) instance runs a primary plus one to three read replicas. If the primary becomes unreachable, KiviDB promotes a replica…
A high-availability (Pro) instance runs a primary plus one to three read replicas. If the primary becomes unreachable, KiviDB promotes a replica automatically and repoints the endpoint at it — your application keeps using the same hostname.
When you need it
Choose Pro when an outage of a few minutes is not acceptable, or when you want to spread read traffic across more than one node.
A single-node instance is not unsafe — every write is persisted to disk, and a lost machine is recovered onto a replacement with its data intact. But that recovery takes minutes rather than seconds, and it is not automatic in the same way. Pro is what buys you the difference.
Pro starts at 8 GB on Google Cloud
On AWS the smallest Pro instance is 4 GB. On Google Cloud it is 8 GB, and that is hardware rather than pricing: the smallest sustained-performance machine Google offers has two vCPUs, and a build-to-order machine type must also have an even vCPU count. There is no one-vCPU shape to sell you, so the smallest honest Pro instance is the 8 GB one.
Azure has the same floor for the same reason. If you want a 4 GB high-availability instance today, that is AWS.
Create one
In the configurator, choose Pro and set Replicas to 1–3. Everything else works as it does for a single-node instance.
Replicas are spread across the zones of your region. KiviDB asks Google to balance them across zones but to fall back to a zone that has capacity rather than failing — so you normally get zone isolation, and when a zone is short you get a working fleet instead of a stalled one.
What you get
A writer endpoint — your normal hostname. Always points at the current primary, including after a failover.
A read-only endpoint — your-name-ro.cloud.kividb.io. Distributes reads
across replicas. Use it for queries that tolerate data a fraction of a second
behind the primary.
Writes sent to the read-only endpoint are rejected. That is deliberate: it stops an application silently writing to a node whose changes would be overwritten.
Failover
If the primary fails, KiviDB promotes the best-caught-up replica and moves the instance's reserved external address onto it. Your clients reconnect to the same hostname and, in most cases, the same address.
That address move is worth calling out because it differs from Azure. There, a promoted node cannot take over the primary's address and the switch has to go through a DNS record with a 60-second lifetime. On Google Cloud the address itself moves, so clients that cache DNS aggressively are not waiting for a record to expire.
The promoted node is also fully removed from the replica group rather than merely pinned inside it, which means nothing can later scale it away or push a replica's configuration onto the machine now serving your writes.
You will get an email and a dashboard notification when this happens, telling you which node was promoted.
Replicas replicate asynchronously, so a failover can lose writes acknowledged in the moment before the primary died. This is the standard trade for staying available; if you need it not to happen, you need a different consistency model than any Redis-compatible database offers.
Changing replica count
Instance → Overview → Scale replicas. You can go up or down between 1 and 3 without downtime. Removing replicas does not interrupt the primary.
You cannot drop below one replica on Pro. To go back to a single node, create a new Essentials instance and restore a snapshot into it.
Upgrading from single-node
Instance → Overview → Upgrade to Pro, then choose a replica count. The existing node becomes the primary and replicas are built alongside it. Your endpoint does not change.
If your instance is below 8 GB you will be asked to resize it first, for the reason above.
With private connectivity
Private connectivity works with Pro. The Private Service Connect attachment publishes the current primary, and a failover moves it to the promoted node, so a private endpoint keeps reaching the writer without anything to change on your side. See Set up private connectivity.