Skip to main content

What to Back Up

How you take backups — the tool, snapshots, object storage, retention, RPO and RTO targets, restore drills — is your organisation's standard and your decision. OLIVE does not replace it and does not document it.

What OLIVE can tell you is the part you cannot see from outside: where the state actually is, which pieces belong together, and in what order they have to come back. Getting that wrong produces a restore that appears to succeed and is quietly incomplete.

OLIVE's one contribution to the mechanics is a label: it marks the volumes that hold state so your backup tool can select them. What performs the backup is yours to choose.


Where the state is​

StateLives inIf it is lost
Connectivity content — datapoints, connectors, templates, authentication entriesThe tenant's CouchDB volumeThe tenant's work is gone. Recoverable from an OCB export, up to the age of that export
Uploaded files and microfrontend componentsThe tenant's SeaweedFS volumeFiles referenced by the content are missing. Microfrontend components are recoverable from an OCB export
Workflow engine stateThe Temporal PostgreSQL database (CloudNativePG)In-flight workflow executions and their history. Content is unaffected
Secrets — storage credentials, OAuth2 Proxy secret, connectivity authentication dataVaultNothing starts. This is the single most damaging loss in the list
Users, realms, clients, role assignmentsKeycloak's databaseNobody can log in, and tenant onboarding has to be redone
The deployment itself — charts, values, tenant listYour Git repositoryRebuildable only from whatever you kept elsewhere. This is why the deployment is in Git
CertificatesRe-issued by cert-managerNothing, if the issuer still works

Vault and Keycloak are cluster infrastructure rather than part of the OLIVE platform, but they are listed here because an OLIVE restore is not possible without them.

Selecting those volumes​

The couchdb and seaweedfs claims carry a configurable label so a backup tool can find them without anyone maintaining a list of claim names. It defaults to olive-boc/backup: velero — the selector OLIVE's own example schedules use — and global.volumeLabels replaces it with whatever your tool selects on:

global:
volumeLabels:
my-backup-selector/include: 'true'

Set it to {} for no labels at all. It has to be decided before the first install: the labels are part of the StatefulSet's volumeClaimTemplates and cannot be changed afterwards. See Storage Requirements.

Temporal's PostgreSQL is a CloudNativePG cluster and is backed up through the operator rather than by labelling a claim. Vault and Keycloak are not labelled by OLIVE at all — they are not its charts.

The OCB export is not a platform backup

An OLIVE Connectivity Bundle contains the connectivity elements and the microfrontend components. It does not contain Vault, Keycloak, workflow state, cluster objects or the deployment configuration.

It is nonetheless the most useful thing in this list for one reason: it is the only copy that is internally consistent by construction, and the only one that can be restored into a different deployment. Keep exports as well as volume backups, not instead of them.


What has to be consistent with what​

The tenant's content and its files are in two separate volumes, and the content references the files. A backup that captures the two at different moments can produce documents pointing at files that are not there, or files nothing points at.

The same applies across tenants and to Vault: a Vault backup older than a credential rotation will not open the storage that the volume backup contains.

So group your backups deliberately:

  • A tenant's CouchDB and SeaweedFS volumes belong to one backup set, taken as close together as your storage system allows.
  • Vault and Keycloak should be no older than the tenant data they have to unlock.
OLIVE provides no quiesce hook

There is nothing in OLIVE that pauses writes, flushes state and signals "safe to snapshot now". A volume snapshot of a running OLIVE is therefore crash-consistent: it captures the disk as it would look after a power cut, and recovery depends on the data store's own crash handling rather than on any guarantee from OLIVE.

If your recovery requirements call for an application-consistent copy, the OCB export is the mechanism OLIVE provides for it. For anything stricter, ask before you design around it: olive-support@boc-group.com.


Restore order​

Restoring in the wrong order does not fail loudly — it produces services that start, fail to authenticate, and retry.

  1. Vault, unsealed. Nothing else can read its own credentials until this is up.
  2. Keycloak, with its database. Every request is authenticated at the edge.
  3. Tenant volumes — CouchDB and SeaweedFS, as a matched pair per tenant.
  4. The deployment from Git, pointed at the restored infrastructure.
  5. Temporal's database, if in-flight workflow state matters to you; otherwise let it start empty.

Then verify rather than assume: the liveness check exercises CouchDB, file storage and orchestration together every five minutes, so a tenant emitting OLIALIVE1 again is real evidence that the restore worked. See Verify the Deployment.

Reusing tenant names after a partial restore

Vault outlives a namespace. A tenant restored into a cluster that still holds the old tenant's Vault policies and auth roles can onboard halfway and then fail identically on every retry. See Uninstall and Teardown.


Not in scope​

Backup tooling and its operation — which tool you use, CSI volume snapshots, object storage targets, encryption of backups, retention schedules, offsite copies, RPO and RTO targets, and restore drills.

OLIVE has no opinion on any of it. What it provides is the label that says this volume holds state worth keeping, plus the grouping and ordering above. Performing the backup is yours.