Metrics and health
Interpret host, guest, storage, and network measurements.
Metrics report observed infrastructure behavior over time. Infra separates measurements from lifecycle state and from the health conditions derived from those measurements.
Measurement sources
Node agents report host CPU, memory, device, and interface observations. Hypervisors report droplet CPU time and virtual device I/O without requiring a guest agent. Guest agents add measurements such as guest memory and filesystem usage when the image supports them.
Storage services report replica placement, capacity, and block latency. Network services report probes, connection counts, and forwarding observations. The control plane reports request health, worker backlog, and reconciliation.
Use the source identifier in a query result when comparing observations. Metrics query API returns source, unit, resolution, and explicit missing buckets rather than collapsing these sources into one health number.
A stopped guest can have no guest samples while its reserved capacity remains unchanged. Guest memory usage and host memory pressure describe different conditions and should not be substituted for each other.
Units and aggregation
CPU usage is a rate over the selected sample interval. Droplet CPU charts show usage relative to its configured vCPU allocation. Memory charts use bytes with binary display units. Network throughput uses bytes per second unless a chart explicitly labels bits per second. Disk IOPS counts operations per second.
The selected time range controls chart resolution. Aggregated data can hide short spikes visible in a narrower range. Use the chart's aggregation label to distinguish average, maximum, and percentile views.
Missing data appears as a gap. Infra does not replace missing samples with zero. An unreachable source can retain its last observed value with a stale indicator. Check the source timestamp before using a value in a maintenance decision.
Health conditions
Resource health combines named conditions such as agent reachability, storage redundancy, capacity pressure, and service probe results. Open a condition to see its source, last transition, and related operation.
Acknowledging an alert does not change resource health. Resolving the underlying condition clears health after a new valid observation. A maintenance window can suppress notifications without changing the measured state.
Filter and compare
Open Observe → Metrics or a resource's Metrics tab. Filter by project, region, zone, pool, node, or label. Compare the same unit and aggregation across resources. Use operation annotations to correlate migrations, resizes, and agent updates with measurement changes.
Labels are evaluated against the selected resource inventory. A historical chart can include data from a resource whose current labels differ from its labels at the time of the measurement; the chart identifies its resource ID.
Export and retention
Installation administrators configure raw and aggregated retention under Configuration. Retention affects which resolutions remain available for older intervals. Export selected data before its retention expires when an investigation needs a durable copy.
Use dashboards to assemble service views, alerts to evaluate conditions, and logs and journals for event-level investigation.
Raw cadence and steal observations
Supported node and droplet charts use a default raw sampling interval of 5 seconds before aggregation. The collection configuration and the agent interval control the actual cadence. Source outages, collection pressure, and unsupported guest capabilities can leave gaps; the chart does not fabricate samples to maintain a visual cadence.
CPU steal is available only from sources reporting that measurement. Compare its source and unit with configured vCPU and scheduling pressure rather than assuming every CPU-utilization metric includes it. Missing steal capability is not a measured zero.
Use Prometheus export, anomaly rules, and availability reports for distinct external and service-level views.