Object storage

Stores and buckets, what one node and several nodes survive, keys limited per bucket, and connecting over S3.

A store is an S3-compatible server of your own. It holds files rather than rows: uploads, images, exports, anything an application writes once and reads back by name. Any library or tool that speaks S3 can use it.

Like a database, a store is ordered separately from any application and outlives them.

Choosing an engine

Three are offered, and all three speak S3.

MinIO. The widely known one. It is also the only engine whose buckets and limited access keys are managed from the ISOGrid console. If you have no reason to pick another, pick this.

SeaweedFS. Fast with very large numbers of small files.

Garage. Lightweight, built to run as a small cluster.

On SeaweedFS and Garage the platform does not manage buckets or limited keys for you. You work with the store's root keys and an S3 client, or with the engine's own web console.

One node or several

Single node. One node, one disk, and that disk is the only copy. If its machine is down, the store is unreachable until it returns. If the disk is lost, the files are lost. Cheapest, and correct for development or for files you can produce again.

Distributed. Three or five nodes, each with its own disk, and the engine copies every write between them as it happens. A node can be lost without losing files. Three or five, never two or four, for the same reason as a database: the nodes need a majority to agree.

The size you choose is per node, and the usable space of a distributed store is less than nodes times size, because the extra copies are what survive a failure. Expect about half on MinIO and SeaweedFS, and about a third on Garage. The store's details show the raw figure, the usable figure, and how many nodes that engine can lose.

Choose the size with room to grow. The console has no way to enlarge a store's disk afterwards, or to turn a single node into a distributed store.

Price before you create

Sizes are named and priced per region, and each one shows its price before you deploy. Single node and distributed are priced separately, and a distributed store runs three or five nodes of the size you pick.

If a region has no priced size for a topology, a store of that kind cannot be created there. Pick another region or the other topology.

Buckets and access keys

This section applies to MinIO stores, once they are running.

Buckets are private. Nothing in one can be read without a key, or a link signed with a key. A bucket name is 3 to 63 lowercase letters, digits and hyphens, with no dots.

The root keys reach every bucket. Each store has one root key pair. Keep it for administration. It is shown on request rather than on opening the page, and every time someone views it is written to the audit log.

Give each application its own key. An access key is limited to the buckets you name, and is either read only or read and write. It cannot create or delete buckets. The buckets must exist before the key is created.

The secret half of an access key is shown once, when it is created, and cannot be shown again. If you lose it, revoke the key and create another. That is also how a key is rotated.

Revoking takes effect at once. Anything using the key stops working, including links that were already signed with it.

Only the person who created a store, or an administrator, can change it, see its root keys or delete it.

Connecting

A new store is private: it has no address outside the platform. Publish it and it gets an HTTPS address of its own, with a certificate. That address is the endpoint you give to an application or an S3 client. Making it private again takes the address away; publishing later gives the same one back.

An S3 client needs four things:

  • Endpoint: the store's HTTPS address, as shown, with no bucket in it.
  • Region: us-east-1 on MinIO and SeaweedFS, garage on Garage. The region is part of how a request is signed, so a wrong one is a refused request.
  • Access key and secret key: a limited key for an application, the root keys for administration.
  • Path-style addressing: the bucket goes in the path, not in the hostname. Most S3 libraries have a setting for it, and it must be on.

Each store also comes with its engine's web console, at a separate address. On MinIO you sign in with the store's keys. On SeaweedFS and Garage the browser asks for a password, shown with the connection details. The console can be turned off.

Replication is not a backup

The platform does not back up a store, and no store keeps older versions of a file. A file that is deleted or overwritten is gone.

A distributed store protects you from a machine failing. It does not protect you from a mistake: the deletion is copied to every node as faithfully as the upload was. If the files matter, copy them somewhere else on a schedule with an S3 tool, and check once that you can read the copy.

Edge cases worth knowing before you meet them

The store says it is running and requests fail. Open its details and use Re-apply configuration. It creates whatever is missing and corrects whatever is running in the wrong shape. It deletes nothing.

Every request is refused with a signature error. The region or the addressing style is wrong. See Connecting above.

The buckets section says it is not available. The store is not MinIO, or it is not running yet.

You deleted a bucket. Every file in it is gone. No copy is kept.

You deleted the store. Every node and every disk is removed, and the files with them. This cannot be undone.

What to read next