latent-sre/save-toolkit/obs-dashboards

Dashboard design

A dashboard should answer a named reader's next operational question. For an existing dashboard, start with the requested design change and preserve the rest; a panel addition does not require a full dashboard redesign.

Grafana implementation lives in `grafana`. Load that skill when the request needs dashboard search/interpretation, panel JSON, variables, datasource discovery, create/edit/export, folders, permissions, validation tooling, version history, or rollback. It owns those procedures and their bundled helper. This skill supplies design decisions and grants no live-write authority.

Design for a decision

Name each panel's backend per stack-profile and flag panels needing a vendor plugin. Use obs-metrics, obs-logs, or obs-traces for query semantics, and the team's folder, naming, time, and variable conventions via grafana.

For example, "Are order requests failing?" starts with error ratio beside request volume and request latency, compared with a known-normal window for the same population. Drill from the affected route to its errors and traces. "Is position sync current?" instead starts with age of last success against its schedule, then the latest run's outcome and failing-stage logs.

Distinguish no traffic, zero, missing telemetry, and query failure; a blank panel must not look healthy, and color is not proof an alert exists or fires. Never append or vector(0) or map null to zero on an error ratio or its numerator; put the request-rate (denominator) panel beside every ratio. Add a telemetry-present stat whose zero is the alarm: on a Prometheus-scrape target use count(up{job="<job>"} == 1) or on() vector(0) mapped red at 0 (unconfirmed against the team's Mimir labels); on a push (OTLP) target there is no up series, so use the backend or pipeline presence signal for that route instead of inventing one.

Check and hand off

State which of the normal, degraded, and missing-data states were actually inspected; a design or screenshot does not prove current health, query correctness, access, or a successful save. Return the audience/question, panel changes, signal/query requirements, units, comparison windows, missing-data behavior, and design limits.