Databases

Ordering one, what each arrangement survives, and why replication is not a backup.

A managed database is ordered separately from any application, because it outlives them. You will redeploy an application a hundred times against the same database.

The platform creates it, keeps it running, patches it, and hands you the credentials. What is inside it is yours.

Choosing an arrangement

Three, and each adds exactly one capability over the one before.

Single. One copy. The cheapest thing that is still a database. No failover: if the machine underneath it goes, the database is unreachable until it comes back. Correct for development and staging, where an outage is an inconvenience.

Single with pooling. The same one copy, with connection pooling in front. It survives a flood of clients but not the loss of its machine. Choose it when your application opens many short-lived connections — which most web frameworks do — and you are not yet paying for availability.

Highly available. Several copies, one of which is the primary and accepts writes while the others follow it. If the primary fails, another takes over and your application reconnects to the same address. This is the only arrangement that survives losing a machine.

Why the number of copies is odd

Three or five, never two or four. Deciding which copy is in charge needs a majority to agree. An even number can split down the middle and agree on nobody — which makes two copies less available than one, because either failure leaves no majority and the database stops accepting writes to avoid corrupting itself.

Connecting

You are given one address. Use it and nothing else. It stays correct when the primary moves, which is the entire point: your application should never know or care which copy is in charge.

The connection details, password included, stay available in the console: open the database and choose Show connection details. For a database that runs on your own cluster through the ISOGrid Nomad agent, the password is kept in your Vault and the console says where.

Backups are not replication

This deserves its own heading because it is the most expensive misunderstanding in this document.

Replication copies every write to every copy, faithfully and immediately. That includes the write you did not mean. A dropped table, a DELETE with a mistaken condition, a migration that ran against the wrong database — all of it is replicated perfectly to every copy within moments.

Replication protects you from hardware failing. Backups protect you from people and software being wrong. They are different problems and they need different machinery. You need both.

A database instance is backed up every day from the moment it is created (a MariaDB instance every 6 hours), and each backup is kept for 7 days, then deleted. A mini-database has a daily schedule of its own, covering that database only. In the Backups view you can move the hour, back up every 6 or 12 hours instead, keep backups for fewer days, take a backup now, download one, or switch the schedule off. Download the backups you want to keep longer than a week. Check once that a restore works before you need it to.

A MariaDB instance runs on one node that you scale up by choosing a bigger size on its page. It is paid 30 days ahead from your balance; a size change refunds the unused days and charges the new size, and deleting it refunds the unused days.

Edge cases worth knowing before you meet them

The database says it is running and your application cannot connect. An instance is marked running as soon as its parts are created, slightly before one of the copies has been elected primary. For the first half-minute or so this is optimistic. Retry.

Connections are refused under load. A database accepts a fixed number of connections. An application with many copies, each opening its own pool, will exhaust them long before the database is actually busy. Use the pooled arrangement, and lower the pool size in your application — a web application almost never needs the default its framework ships with.

A failover happened and in-flight transactions failed. This is correct. Anything uncommitted when the primary died is gone; a promotion cannot invent the result of a transaction nobody finished. Applications that matter should retry a failed transaction rather than assume it succeeded.

A failover happened and a very recent write is missing. Copies follow the primary with a small delay. If the primary dies in that window, the write it acknowledged may not have reached the copy that took over. The window is short but it is not zero. If your application cannot tolerate that at all, say so before you order — the trade is slower writes for no loss, and it has to be chosen deliberately.

A copy is far behind. Usually a long-running query on that copy, or a burst of writes larger than the link can carry. It catches up on its own. A copy that is too far behind is not promoted during a failover, because promoting it would discard more than the alternative.

The disk filled up. Everything stops, including the ability to delete rows — which needs to write. Watch the size; do not wait for it.

You deleted the database. It is gone, and so is everything in it. The confirmation is deliberately tedious.

What to read next

Seeing how it is doing

Every PostgreSQL database has an Insights view, refreshed every few seconds while it is open.

For a database instance - a server of your own - it covers the whole server: connections against the limit, transactions per second, how much is read from memory rather than disk, checkpoints and the write-ahead log, replication and its lag, what is running and who is blocked by whom, the slowest statements, and the largest tables and indexes of any one of its databases. An index that has never been used and a table that needs a vacuum are pointed out.

For a mini-database it covers that one database and nothing else: its size, its connections, its transactions, its tables and indexes, and what is running in it. A mini-database shares its server with other tenants, so nothing about the server or any other database on it is read - not their names, not their traffic, not the server's settings.

Slowest statements are collected on instances created from October 2026 on.

Extensions

pgvector (vector similarity search), PostGIS (geographic data), pg_trgm, unaccent, citext, hstore, ltree and a dozen more are switched on from the database's Extensions view. The platform creates the extension for you: most of them need a superuser, which your own role is not.

Switching one off fails, with PostgreSQL's own explanation, while a column or an index still uses it. Extensions that reach out of the database or run code as the superuser are not offered.

An instance created before an extension arrived lists it as not on the server yet: redeploy the instance to move it onto the current image.