JWT or API-key access, tenant isolation, policy and approval surfaces.
Architecture and security
Keep control centralized and data processing local.
Ark separates orchestration from execution. Teams manage policy, access, and jobs centrally while an outbound-only agent processes row-level data inside the customer network.
Agent protocol gRPC v3
- Connection
- Agent-initiated, outbound only
- Identity
- Per-agent X.509 certificate
- Row processing
- Inside the customer network
- Source access
- Read-only user or replica
✓ Raw data never leaves your VPC
Trust boundary
A clear separation between control and data planes.
The control plane knows what should run. The agent knows how to execute it beside the source. Raw rows stay on the customer side of the boundary.
01 / Control plane Control Plane Managed SaaS or self-hosted
WEBCLIGO SDKJS SDKCI
Persists jobs, leases work, and coordinates connected agents.
Tenant metadata, configuration versions, audit events, health, and artifact references.
Agent initiates ←
mTLS · gRPC v3 ↔ Long-lived bidirectional stream for heartbeat, assignment, lease, progress, and completion.02 / Data plane Customer VPC / private network Customer-owned runtime and network
Maintains the mTLS session, capabilities, lease, and job lifecycle.
Runs classification, masking, relational subsetting, synthesis, and test environment actions locally.
PostgreSQL / MySQL source Test target / ephemeral DB Customer object storage Optional local Ollama
Request-to-result flow
A job remains governed from request through completion.
The control plane owns authorization and state. The agent owns database access and execution.
- 01 Client Request and authorize
A user or CI runner submits a job through REST using JWT or an API key.
- 02 Ark Control Persist and validate
Ark Control applies tenant and RBAC checks, resolves the approved configuration, and records the pending job in ark-meta.
- 03 Ark Control Assign and lease
The connected agent receives a job assignment over the bidirectional gRPC stream and accepts a bounded lease.
- 04 Ark Agent / VPC Fetch the execution package
The agent downloads the authorized job package over gRPC and resolves source credentials locally.
- 05 Ark Agent / VPC Execute beside the data
The agent reads the source with least privilege, applies policy, and writes only to approved targets or customer storage.
- 06 Ark Control Report and finalize
Progress, derived JSON artifacts, checksums, target readiness, and completion status return to Ark Control for audit and delivery.
May cross the boundary
Operational metadata
- Schema names and column types
- Classification labels and confidence
- Job state, health, and execution logs
- Policy references and audit events
Never required outside the VPC
Sensitive data and secrets
- Raw production rows
- Unmasked PII and free-text content
- Production source database credentials
- Prepared dumps in customer storage
Boundary note: production source credentials and raw rows stay local. Approved governance JSON, derived metrics, artifact references, and generated test-environment connection details may return to Ark Control when the workflow requires them.
Network and machine identity
One outbound channel with a verifiable agent identity.
Protocol v3 removes the legacy secondary HTTP agent path. Packages, artifacts, callbacks, and the control stream use gRPC with mTLS.
X.509 Unique certificate per agent
O = tenant UUID CN = agent UUIDARK CONTROL Mutual TLS verification
Ark Control verifies the issuing CA, tenant and agent identity, plus active or revoked certificate state.
- Inbound access to Agent
- None required
- Agent egress
:8081/TCP· TLS endpoint for Ark Control gRPC v3- Source database
- Customer-local PostgreSQL/MySQL access; a SELECT-only user or replica is recommended
- Targets and storage
- Customer-approved network access to test databases, Docker targets, or object storage
- Production source secrets
- Resolved by the Agent from local environment or secret management; only a credential reference is registered centrally
Deployment options
Choose where the control plane runs.
Both options preserve the same principle: row-level work stays inside the customer environment.
Managed Ark Control
Use Subsetra's managed control plane while the data plane remains in your VPC.
- The same In-VPC execution boundary
- Faster control-plane onboarding
- Managed upgrades and availability
Self-hosted Ark Control
Run control and data planes inside infrastructure owned by your organization.
- The same In-VPC execution boundary
- Private control-plane network placement
- Customer-owned operations and upgrades
Minimum network model
Designed for restrictive environments.
Ark avoids inbound access to the agent and supports least-privilege connectivity to sources and targets.
- 01Read-only source access
- 02Outbound-only agent
- 03A Linux, Docker, or Kubernetes runtime for the agent
- 04Customer-approved access from the agent to target storage or databases
Free product demonstration