Start with business operations
Choose operations that matter: sign-in, opening an order, posting a document, payroll, period close, exchange and scheduled jobs. Define expected time, acceptable error rate, measurement window and owner for each.
Do not average everything into one number. Fast navigation can hide rare but critical period-close delays. Separate interactive work, background jobs and integrations.
Use APDEX deliberately
APDEX groups executions as satisfied, tolerating and frustrated relative to a target. The threshold must reflect user expectation rather than a convenient reporting value.
Keep distributions and percentiles beside the index. APDEX alone cannot explain cause and may hide different workload modes. Compare it by unit, database, version, operation and hour.
Collect data by layer
The platform provides key-operation metrics, registration log and technology log. The database adds queries, locks, waits, plans and replication state. The operating and virtualization layers contribute CPU, memory, storage, network and resource contention.
Correlate events through consistent time, database name, cluster node, session and a correlation identifier where available. Enable technology-log events selectively with retention limits: collecting everything continuously creates volume, noise and data-exposure risk.
Alert on impact
An alert should state the affected operation, user population, duration and preceding change. High CPU without a service-objective breach may require observation, but is not automatically an incident.
Use multi-signal rules: APDEX degradation plus error or queue growth, or delayed exchange plus missing control event. Define confirmation windows, deduplication and escalation.
Preserve diagnostic context
On trigger, capture platform and configuration versions, background activity, application release, recent changes, database state and infrastructure context. Mask personal and financial data and restrict access.
After an incident, update the control scenario and baseline. A fix is proven by restored SLO under comparable load, not merely by the absence of another complaint.
A useful minimum
Begin with five to seven critical operations, a dependency map, time synchronisation, APDEX and percentile dashboards, a controlled technology log and one diagnostic runbook. This is more useful than hundreds of ownerless metrics disconnected from business outcomes.

