Health you can read at a glance
Agents probe every member and report conditions such as whether the cluster can read, can write and has quorum, each tracked separately. A condition changes only after two consecutive observations, and evidence that is too old turns it to “unknown” instead of pretending everything is fine. The same health report drives the console, the alerts and an AI assistant’s diagnosis.
Metrics
Agents collect engine and host metrics every 15 seconds into a time-series database built into the control plane, which answers PromQL queries. Each engine has a curated set of charts; you can compare with the previous period and see failovers, configuration changes and your own deploy markers as annotations on every chart.
- Metrics are kept for 15 days or 40 GB, whichever comes first.
- A Prometheus-compatible endpoint lets Grafana read them directly, and ready-made dashboards ship for every engine.
- Large fleets can send metrics to Prometheus, VictoriaMetrics or Mimir instead; that becomes necessary beyond about 500 servers or 30 days of history.
- Heavy queries are limited so they can never delay alert evaluation.
Logs and events
Logs are fetched from the servers when you look at them, with counts per level for the last 24 hours. Multi-line entries are joined, levels are normalised across engines, and secrets and literal values in slow-query lines are masked before they leave the server. A 90-day event timeline per cluster records failovers, configuration changes, operations and alerts in order.
Alert rules with runbooks
About 100 alert rules ship with ElastiCluster, each with a concrete threshold, a duration, a severity and a runbook that says what to do. Examples include a cluster without a primary, replication lag, failing log archiving, a disk filling up, an overdue restore drill and an agent that stopped reporting. A rule without a runbook fails the build.
- Override thresholds per project, environment or cluster.
- Acknowledge an alert until it resolves, or silence matching alerts for a while, with undo.
- When more than one server in ten drops at once, a single alert about the outage replaces dozens of individual ones.
