---
title: What disaster recovery should a self-hosted AI governance platform have?
url: https://opencontroller.lyzr.com/blog/disaster-recovery-for-a-self-hosted-ai-governance-platform
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What disaster recovery should a self-hosted AI governance platform have?

**Plan recovery per component, not for the platform as a whole. The gateway in the request path needs the shortest recovery time because agents stop or run ungoverned when it fails. The registry, policies and audit data need the smallest data loss. Most teams start with multi-zone high availability plus data-only replication to a second region, then add a warm standby once the platform carries production traffic.**

Security and infrastructure teams ask this early in a self-hosted rollout, usually as two questions: what survives a zone failure, and what survives a region failure? They have different answers.

## What is the difference between a zone failure and a region failure?

- **Zone failure:** one data centre (availability zone) in a cloud region goes down. If the platform runs replicas across at least two zones with zone-redundant storage and databases, it should keep running with no manual action.
- **Region failure:** a whole region is unavailable. Nothing inside that region helps. Recovery needs data and, eventually, compute in another region.

Zone resilience is mostly a deployment setting. Region resilience is a design and budget decision.

## Which components need which targets?

Recovery time objective (RTO) is how long you can be down. Recovery point objective (RPO) is how much data you can lose.

| Component | Why it matters | Suggested RTO | Suggested RPO |
|---|---|---|---|
| LLM or agent gateway | In the request path; outage stops agents or bypasses policy | Minutes | Not applicable (stateless), but config must replicate |
| Policy and configuration store | Defines what is allowed; stale policy is a risk | Under 1 hour | Minutes |
| Agent and MCP registry | Source of truth for ownership and approvals | Hours | Minutes to 1 hour |
| Audit and trace store | Evidence for regulators and incidents | Hours to a day | Near zero for audit logs |
| Analytics and dashboards | Useful, not critical | A day | Hours |
| Discovery jobs | Can re-run after recovery | A day | Not applicable |

These are starting points. Your own targets should match how critical the governed agents are.

## Which DR strategy should you pick?

The common cloud DR strategies, from cheapest to fastest, are backup and restore, pilot light, warm standby and multi-site active/active.

1. **Backup and restore.** Snapshots copied to another region. Cheapest. Recovery takes hours because you rebuild the stack first.
2. **Data-only replication (pilot light).** Databases and object storage replicate continuously to a second region; compute is defined in code but not running. A good first step: low cost, low data loss, recovery in tens of minutes to hours.
3. **Warm standby.** A scaled-down copy runs in the second region. Recovery is mostly scaling up and switching DNS.
4. **Active/active.** Both regions serve traffic. Fastest, most expensive, and hardest to keep policy consistent.

Many teams pilot data-only replication first, measure real recovery time in a drill, then decide whether the gateway needs a warm standby.

## What happens to agent traffic while the platform is down?

Decide this before the outage, in writing:

- **Fail closed:** agents stop calling models until the gateway is back. Safe for high-risk agents, disruptive for everything else.
- **Fail open:** agents call providers directly. Keeps business running, but without policy, redaction or cost attribution. Requires credentials that only work in an emergency.
- **Per-agent choice:** high-risk agents fail closed, low-risk agents fail open. Usually the right answer. See [fail-open vs fail-closed AI guardrails](/blog/fail-open-vs-fail-closed-ai-guardrails).

## What should your DR checklist include?

1. Replicas spread across at least two zones for every stateful and stateless component.
2. Zone-redundant databases and storage.
3. Continuous replication of policy, registry and audit data to a second region.
4. Infrastructure as code that can stand up compute in the second region.
5. Secrets and certificates available in the second region.
6. Documented DNS or endpoint failover for the gateway.
7. A written fail-open or fail-closed decision per agent risk tier.
8. A DR drill at least twice a year, with measured RTO and RPO.
9. Runbooks owned by a named team, including who decides to fail over.

## How Opencontroller does this

Opencontroller runs as Lyzr SaaS or on customer-controlled infrastructure. In a self-hosted deployment, Lyzr engineers embed with your team during integration, and recovery design follows your own cloud's zone and region architecture. Per-agent Block and Quarantine controls and policy as code let you set different behaviour for different risk tiers. See [LLM gateway high availability](/blog/llm-gateway-high-availability), the [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is multi-zone deployment enough for DR?**
It covers zone failures. It does not cover a region outage, which needs data replicated to, and a way to run compute in, a second region.

**Should the audit store be active/active?**
Rarely. Continuous replication with near-zero data loss usually satisfies evidence needs without the cost of active/active.

**How often should we test DR?**
At least twice a year, with measured recovery times compared against your RTO and RPO targets.

## Key facts

Self-hosted AI governance DR should be planned per component. Gateway (request path): RTO minutes. Policy store: RTO under 1 hour, RPO minutes. Registry: RTO hours. Audit store: RPO near zero. Zone failure is handled by multi-zone replicas and zone-redundant storage; region failure needs cross-region replication. Strategies: backup and restore, pilot light (data-only replication), warm standby, active/active. Decide fail-open vs fail-closed per agent risk tier before an outage. Run DR drills at least twice a year.

## Related

- [How do you make an LLM gateway highly available?](https://opencontroller.lyzr.com/blog/llm-gateway-high-availability.md)
- [Should AI guardrails fail open or fail closed?](https://opencontroller.lyzr.com/blog/fail-open-vs-fail-closed-ai-guardrails.md)
- [Should you run an AI governance platform as SaaS or self-hosted?](https://opencontroller.lyzr.com/blog/saas-vs-self-hosted-ai-governance-platform.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
