Original practice scenario: A network simulator generates thousands of alarms after one upstream link fails. Design incident grouping without suppressing unrelated failures.
Read model answer +
Group alarms using topology and time evidence while retaining raw events. Correlation should create a reviewable hypothesis, not assert that every nearby alarm has the same cause.
Work through it
Version the dependency graph and capture alarm identities, onset times, recovery signals and confidence in their device mappings.
Form candidate incidents using affected paths and bounded time windows, preserving independent symptoms and allowing groups to split when evidence conflicts.
Test simultaneous unrelated failures, stale topology, delayed alarms and recovery flapping; measure missed incidents and unnecessary duplicate notifications separately.
Tradeoffs to discuss
Aggressive grouping reduces noise but can hide distinct failures; conservative grouping increases workload while preserving more uncertainty.
Common pitfalls
Do not discard raw alarms after attaching them to an incident.
Temporal proximity alone is insufficient evidence of a shared root cause.
Original practice scenario: A hypothetical operator forecasts traffic after adding new cells. Historical growth mixes customer demand with changes in where traffic is measured.
Read model answer +
Distinguish underlying demand growth from redistribution across a changing network. A new cell can shift existing traffic rather than create new customer activity.
Work through it
Maintain effective-dated coverage and routing metadata, and construct stable regional aggregates alongside cell-level series to identify redistributed volume.
Backtest using only information known at each planning date, including planned deployments as uncertain scenarios rather than confirmed future observations.
Evaluate peak demand, intervals and capacity shortfall by region, comparing simple baselines and stress scenarios with delayed or cancelled deployments.
Tradeoffs to discuss
Regional models are more stable across topology changes but can miss local bottlenecks; cell models need explicit treatment of network reconfiguration.
Common pitfalls
Do not compare new-cell traffic with a fictional zero-demand history.
Forecast accuracy and the cost of underprovisioning are separate evaluation targets.
