YugabyteDB metrics analysis

Help an operator pull the right metrics from a YugabyteDB universe and turn them into a diagnosis. This skill is about observability-driven triage: given a symptom (or no symptom — "is this universe healthy?"), choose a metrics source, locate the relevant series by name, query them at the right aggregation, and read the trends that confirm or rule out a specific class of problem.

Output discipline — advisory only; no changes without explicit approval

This skill is read-only: it queries metrics (PromQL, the YBA metrics proxy, /prometheus-metrics scrapes) and interprets them. The default deliverable is succinct guidance — a diagnosis and recommended remediation for the user to review, not changes you apply. Every remediation here (scale out, resize, drop unused tables/indexes, edit GFlags/runtime config, raise tablet/overhead limits, trigger compaction, restart) is routed to yba-api or the operator and requires the user's explicit approval — never apply one yourself. Present the exact action and its effect; let the user authorize it.

It complements the yba-api skill (control-plane automation): for raw Prometheus connection and PromQL plumbing against a YBA-bundled Prometheus, the yba-api Prometheus reference is the canonical low-level guide — this skill links to it rather than repeating it.

Related skills:

• Whole-universe performance assessment (open-ended "assess this cluster for any performance issues" — runs this metrics track and the SQL track, then synthesises) → start from yb-performance-assessment. Use it when the scope is the whole universe rather than a specific metrics symptom.
• SQL-layer query analysis (slow queries, pg_stat_statements, ASH, locks, query tuning, pre-production assessment) → yb-query-analysis
• YugabyteDB Aeon REST API (cluster scaling, backups, allowlists, metrics API) → aeon-api

Scope: where this skill applies (and where it doesn't)

This skill targets self-managed YugabyteDB and YugabyteDB Anywhere (YBA)–managed universes — anywhere you can reach a Prometheus (bundled or standalone) or the nodes' /prometheus-metrics endpoints directly. YugabyteDB exposes on the order of ~2,000 metrics there; the playbooks here assume you can query that full surface.