id: database-latency-triage name: Database Latency Triage version: v2 description: Diagnose database latency, pool saturation, lock contention, and dependency pressure without modifying data. triggers: [database latency, slow query, connection pool, lock wait, deadlock] services: [mysql, postgres, database] resourceTypes: [service, pod] categories: [database, dependency] applicableTasks: [QUERY_LOGS, QUERY_METRICS] tags: [database, latency, connection, lock, dependency] maxRisk: LOW toolWhitelist:

• knowledge.searchSop
• kubernetes.describeResource
• kubernetes.getPods
• kubernetes.queryLogs
• kubernetes.queryEvents
• kubernetes.queryPodLogs
• kubernetes.queryMetricsContext
• prometheus.queryRange
• prometheus.instantQuery
• prometheus.rangeQuery
• alerts.getActiveAlerts
• changes.getRecentChanges
• alertmanager.listAlerts

Database Latency Triage

Diagnostic objective

Distinguish database saturation from application pool, retry, network, and transaction behavior using current evidence.

Required live evidence

Collect request latency and errors, pool wait/utilization, active connections, query latency, lock waits, long transactions, replication lag, and the change timeline. Compare an affected service with a healthy peer.

Decision branches

• Pool wait rises while database latency is normal: inspect client limits, leaks, and retry concurrency.
• Query and lock latency rise together: identify the transaction pattern and owner without killing sessions.
• Replication lag rises: separate replica read pressure, network delay, and primary write saturation.
• Application retries lead the database spike: treat retry amplification as a contributing cause.

Safe next step

Preserve query identifiers and timestamps, limit conclusions to verified signals, and use the dependency owner runbook for approved mitigation.

Prohibited actions

Do not clean data, kill sessions, alter schema, fail over, or change database configuration from this Skill.

Recovery and escalation

Require latency, errors, pool wait, locks, and lag to remain normal across two windows. Escalate for data integrity risk, primary unavailability, or sustained connection exhaustion.