Skip to content

System health

Platform › Infrastructure › System health shows whether every service the platform depends on is answering right now, grouped by what it does, and how often each one answered over the last 30 days. It is the first page to open when something in the console or at the gateway stops working.

Platform administrators only.

  1. The header, System health, with “Live status of every platform dependency the control plane relies on.” A Live · every 15s tag and Refresh sit beside it.

  2. The roll-up banner:

    Banner Meaning
    All systems operational Every configured dependency is responding.
    Degraded One or more components are unreachable, none of them core.
    Service outage A core dependency is unreachable.
  3. Four tiles: Healthy, Down, Not configured and Components.

  4. The components, grouped by category. Each tile shows the component’s name, a Core tag when it is one, its state, the detail of the last failure, and a line such as “99.8% · 30 days” with a small trend of the share of each day it answered.

State Meaning
Healthy It answered the last probe.
Unreachable It is configured and did not answer. The tile shows what the probe got.
Optional — not enabled The installation does not run it. This is not a fault, and it does not count against the roll-up.
Category Components
Core The control plane, the platform database and the cache are core: any of them down is an outage. The metering ingest is in this group but is not core: while it is down, inference still serves and metering catches up when it returns.
Data & storage Object storage, where uploaded documents are kept, and the vector databases tenants’ knowledge lives in.
AI services The guardrails service, the reranker, the document converter, the model catalog source (Hugging Face, or your bastion’s mirror of it), and model serving: each ready deployment’s health endpoint, naming any that does not answer.
Observability The tracing service, when tracing is on.
Gateway & identity The gateway applications call models through, and the identity provider sign-in depends on.
Model providers One tile per enabled cloud provider connection, probed by listing its models. A probe costs no tokens.

A component the installation does not run appears as Optional — not enabled. Turning optional components on is an installation setting: see Configuration.

  • Refresh reads the components now instead of waiting for the next 15-second refresh.
  • Read an Unreachable tile’s detail for what the probe got back, such as a refused connection or an HTTP status.

Olympusx AI samples every component once a minute, whether or not anyone has the page open, and keeps 30 days of samples for the daily figures.

  • A component that fails three samples in a row raises a platform notification that it is down, linking to this page.
  • When it answers again, a second notification says it recovered. A component nobody was told was down raises no recovery notice.
  • An optional component that is not enabled never raises one.
Limit Value
Page refresh Every 15 seconds
Background sampling Once a minute
History kept for the daily figures 30 days
Failed samples before a down notification 3 in a row
Component What stops
Control plane Everything: nothing routes and the console is blank.
Platform database The console and every management action.
Cache Rate limits and budget checks. See Monitoring and alerts for how the gateway behaves meanwhile.
Metering ingest Nothing visible to callers. Inference is served, and usage is recorded when it returns.
Gateway All inference by API key. The console still works, which makes this easy to misread.
Identity provider Signing in.
Guardrails service Every tenant whose policy enforces guardrails, unless the platform is set to fail open. See Settings.
Document converter, reranker Document ingestion and reranking. Chat is unaffected.
A model provider Calls to that provider’s models.