For the complete documentation index, see llms.txt.
Skip to main content
Version: 8.10 (unreleased)

Physical Tenant isolation model

info

This is the detailed technical documentation for Physical Tenants. For an overview and key concepts, see Physical Tenants.

Physical Tenants provide strong isolation within a single orchestration cluster. This page assumes one orchestration cluster with multiple Physical Tenants. Multi-region and multi-cluster topologies are separate topics.

Isolation model

A Physical Tenant is an isolated execution unit inside one orchestration cluster.

LayerIsolation modelShared or isolated
Primary storageDedicated Raft groups per Physical Tenant. A single tenant can span multiple brokers.Isolated
BrokersBrokers are co-located and can host more than one Physical Tenant.Shared infrastructure
GatewaysGateways route requests to the targeted tenant.Shared
Secondary storageUse a tenant-specific schema, index prefix, or separate backend, depending on the storage type.Isolated
Document storeUse a tenant-specific bucket, container, or subpath. The exact convention depends on the cloud provider.Isolated

Architecture

The diagram shows one orchestration cluster boundary with shared control-plane components and tenant-specific execution and storage boundaries.

API routing

Use tenant-scoped routes for tenant-specific requests:

  • REST: /physical-tenants/{physicalTenantId}/v2/...
  • gRPC: Camunda-Physical-Tenant
  • Default tenant compatibility: if the gRPC header is omitted, the request routes to the default Physical Tenant

Use cluster-scoped routes for operations that apply to the whole cluster:

  • Cluster-wide REST: /cluster/v2/...

What is not isolated in 8.10

  • Gateways are shared between tenants, so a saturated gateway can still affect multiple tenants.
  • Brokers are co-located and shared infrastructure remains part of the deployment.
  • Full performance isolation is out of scope for the first version.
  • Future versions may reduce sharing further, for example through more isolated actor-thread or runtime placement, but that is not part of 8.10.

Storage validation

Configuration validation should fail fast when two tenants point to the same backend location or another unsupported path. For document store, the final naming convention depends on the provider, so validate uniqueness at startup rather than relying on a hard-coded path format.

Readiness

The intended model is that readiness remains broker-scoped, not tenant-scoped. If one physical tenant loses access to its secondary storage, that should not fail readiness for the entire cluster.

For the storage backends used by tenant-scoped data, see secondary storage and document handling configuration.