Skip to main content

Bring Your Own Cloud (BYOC)

A BYOC deployment runs inside your own environment, on private networking with no public inbound, in one of two modes: connected or airgapped. Most of this page describes connected BYOC, whose shape is the same on AWS and Azure. Each tab shows the concrete services. For how BYOC compares to SaaS, see Deployments.

note

The trust posture, in short:

  • The application, data and identity stay in your account. Inference with an in-account model stays there too.
  • Data leaves only on the egress paths you choose.
  • There is no public inbound.
  • Autoheal holds no standing credentials to your infrastructure.

In connected BYOC, operator access is optional, granted by you and revocable at any time. In airgapped BYOC there is no operator access.

Connected and airgapped​

BYOC runs in one of two modes. The topology on this page is connected BYOC. The platform runs on your EKS or AKS cluster. Release images are pulled into your account for each upgrade.

Airgapped BYOC is for environments with no path to the internet or to Autoheal. The platform runs on Docker Compose on hosts you provision. Every infrastructure and application image arrives in a versioned bundle with checksums, which your team imports. Nothing is pulled from a registry at runtime. Autoheal has no operator access. Each upgrade is a new bundle that your team imports on its own schedule. LLM inference goes to a model endpoint reachable from the isolated network, which you configure. See Sovereignty for how the two modes compare.


The concrete services behind each tier:

  • Networking — a VPC across three Availability Zones, private subnets only, no public inbound. Users reach it over VPN or AWS Direct Connect through a Route 53 private hosted zone (ACM certificate). Ingress terminates at an internal Application Load Balancer (443). Egress is outbound-only over VPC endpoints / PrivateLink where available.
  • Compute — Amazon EKS with a private API endpoint, plus a separate Runner (EC2 Auto Scaling group) for sandboxed execution. The runner has no inbound: it connects outbound to the in-cluster Temporal frontend over an internal NLB (7233), polling for work.
  • AI models & inference — in-account Amazon Bedrock (via PrivateLink) keeps inference in your account. You can also choose a direct provider.
  • Data — Aurora PostgreSQL, ElastiCache for Redis, Amazon S3, AWS Secrets Manager, all in your account. All are encrypted at rest with your KMS keys. Secrets are read through EKS Pod Identity / IRSA.

What leaves your environment​

The entire platform runs in your account, so by default nothing leaves your environment. With an in-account model and in-network integrations, no data crosses your boundary at all. Data leaves only on paths you deliberately choose:

Where it goesLeaves your environment?
Platform services + all data at restYour accountNo — never
LLM inference — in-account modelYour account, over a private endpointNo
LLM inference — direct providerThe provider you chooseOnly if you pick one instead of an in-account model
Integration reads — in-network toolsInside your networkNo
Integration reads — external SaaS toolsThat vendorOnly for the SaaS tools you connect
Inbound — alerts and user sign-inInto the platform, on private networkingNo public inbound

There are two egress paths: a direct LLM provider, and reads to external SaaS integrations. Both are choices you make, not something Autoheal does on its own. See Data residency. Inbound traffic is limited to two sources: your users signing in over SSO, and your monitoring posting alerts to an authenticated webhook. In BYOC, both use private networking, with no public inbound.

Access and identity (SSO)​

  • Connect your SAML 2.0 identity provider for single sign-on (Okta, Entra ID, PingOne, OneLogin, Google Workspace, or an in-house IdP). Your team signs in with existing corporate credentials. OIDC federation is not currently supported. Customer SSO is SAML 2.0.
  • Every request is authenticated at the gateway, and access is role-based, keyed on the user's email. Autoheal never receives passwords or MFA factors.
  • Role bindings key on email, so changing a user's email in your identity provider re-keys their access. Plan email changes as a re-grant, not a rename.

Data residency​

  • All data resides in your infrastructure — databases, cache, object storage, and secrets, all in your account and region.
  • Encrypted in transit (TLS) and at rest, with your own keys (KMS or Key Vault) where policy requires.
  • Nothing is replicated to another region unless you configure it. The only data that leaves is the LLM and integration traffic you configure.
  • Choosing an in-account model (Bedrock, Azure OpenAI / AI Foundry) keeps inference inside your account over a private endpoint. Choosing a direct provider (OpenAI, Anthropic) is a deliberate outbound egress exception that you opt into. On this path, prompt and context data leaves your account for a model provider's public endpoint.

Backup and restore​

All durable state lives in your account's managed services. Backup and restore run under your organization's own retention and DR policy, with native cloud tooling.

  • PostgreSQL is the system of record — backed up with managed automated backups, point-in-time recovery, and snapshots, encrypted with your keys.
  • Object storage (attachments, artifacts) uses versioning, with optional cross-region replication.
  • Redis is a transient streaming and cache tier. Durable events are persisted to PostgreSQL, so Redis needs no backup of its own.
  • The application runtime is stateless and redeployed from version-pinned releases, so the running services hold no state to back up.

To recover, your platform team restores PostgreSQL to a point in time (and any object versions) from your managed-service backups. Then the application is redeployed to reconnect to the restored data. Your team does the redeploy. In connected BYOC, Autoheal does it instead if you grant the scoped install role. Because backups are encrypted with your keys, protecting those keys is what ultimately protects your data. Detailed recovery-point/recovery-time targets and DR procedures are covered during onboarding.

Deployment and upgrades​

Autoheal holds no standing credentials to your infrastructure. In connected BYOC there are two ways to run installs and upgrades, and you choose:

  • You run it (recommended). Run the install and every upgrade from your own CI/CD with your own credentials. Autoheal never touches your account.
  • Autoheal runs it during onboarding. You create a scoped, least-privilege IAM role (with an external ID). Autoheal assumes this role only for the install window. That principal gets an EKS access entry (cluster-admin on the app's cluster), so it can create the application's Kubernetes resources. You can delete the role afterward. The cluster-admin grant is scoped to the install window and the app's resources, not a standing operator login.

A few properties hold in both install options:

  • Runner hosts have no inbound SSH. On AWS, they are managed through SSM. There is no bastion or jump box, and no open inbound path for operators.
  • You control the upgrade window. Upgrades are staged under change control.
  • The application layer moves independently of your cloud infrastructure, so a routine upgrade never touches your network or data services.

In airgapped BYOC, Autoheal has no access at all. Each release arrives as a versioned bundle of every image, with checksums. Your team verifies the checksums, imports the bundle on the hosts, and starts the new version on its own schedule.

For how an in-flight investigation survives a runner being replaced during an upgrade or scale-in, see Durability and failure handling.