Skip to content
Karakoram

0014 — Built-in agent metrics, Prometheus output and UI charts

Status: Accepted

Context

Section titled “Context”

Per-member visibility requires MongoDB application telemetry in addition to Zerops platform metrics. OpenEverest’s PSMDB monitoring integration uses an independently operated PMM server and member-side PMM clients, not a standalone exporter endpoint.

Decision

Section titled “Decision”

Collect MongoDB metrics in the node agent through its local driver using the purpose-specific clusterMonitor account. Expose Prometheus text through authenticated private agent /v1/metrics. The control plane scrapes/aggregates member samples, retains bounded UI time series and serves:

  • /api/v1/clusters/{c}/metrics?range= for the UI dashboard;
  • /metrics as a Prometheus-compatible endpoint for external monitoring, protected by Bearer/API-token authentication and monitoring RBAC.

Show member role/health, replication lag, uptime and resource/disk observations alongside charted metrics. Include stable cluster/member labels and honor each cluster’s monitoring setting. Apply per-object authorization to aggregation so a permitted scrape cannot expose unrelated clusters. Unavailable or stale data is marked as such, not reported as a healthy zero. Monitoring failures do not redefine database readiness or substitute for direct engine health checks.

Defer PMM integration; it is not required or shipped in the initial supported monitoring path. The agent exporter, public scrape endpoint and UI charts are required, not deferred along with PMM.

Consequences

Section titled “Consequences”

A self-contained installation satisfies dashboard/exporter needs without another server or PMM client distribution. Operators can integrate their own Prometheus-compatible tooling. Built-in short-term samples are not an HA, long-retention monitoring platform; control-plane downtime interrupts aggregation while databases keep running. Rich PMM query analytics and its wider workflows are not claimed. Metric cardinality, retention and stale-sample handling need explicit bounds.

Alternatives considered

Section titled “Alternatives considered”
  • PMM server plus clients: richer analysis and descriptive OpenEverest parity, but adds service credentials, deployment cost and another operational dependency.
  • External exporter only: usable Prometheus integration, but leaves the required self-contained UI dashboard disconnected.
  • Platform metrics alone: cannot supply replica-set/application-level health.