InterviewsVector
AdvancedStaffPerformance, Observability & Capacity55 minutesPro answer

SDV-049

Diagnose a p99 regression that every average says does not exist

Localize tail amplification across cohorts, fan-out, queues, pauses, and dependencies while avoiding coordinated omission and misleading percentiles.

PerformanceObservabilityCapacity PlanningMulti Tenancy

Interview prompt

Problem context

Median latency is unchanged, CPU averages are healthy, and dependencies meet their own p99s, yet 1% of enterprise requests exceed eight seconds after a rollout. Requests fan out, workloads vary by tenant, and load tests do not reproduce production tails.

Skills being evaluated

tail-latency analysismeasurement correctnesscohort debuggingperformance experimentation

The full reasoning guide is part of Pro

The scenario and evaluation focus above remain public. Pro unlocks the structured answer, trade-off analysis, follow-up probes, common weak answers, rubric, related reasoning, and any architecture diagram.

Sign in to continue

What the full guide covers

Clarify the decision
Establish scale assumptions
Functional and non-functional requirements
High-level architecture
Data model and flow
Consistency and transaction boundaries
Failure modes and recovery
Security and privacy
Observability and SLOs
Capacity and cost
Alternatives and trade-offs
Evolution and migration
What Staff and Principal candidates should emphasize

Related architecture decisions

Staff · 55 min

Design observability that stays useful under cardinality and cost pressure

Choose metrics, traces, logs, exemplars, retention, and tenant debugging paths without letting unbounded labels bankrupt or disable telemetry.

Staff · 55 min

Define launch capacity when every demand forecast is uncertain

Convert demand ranges and workload units into tested capacity, admission policy, ramp gates, and business contingency for a high-visibility launch.

Principal · 60 min

Design performance isolation for mixed interactive and analytical workloads

Protect latency SLOs from scans, compaction, and background work using admission, pools, replicas, workload contracts, and resource accounting.

Senior · 50 min

Define SLOs and error-budget policy for a product with unequal journeys

Choose customer-centered indicators, windows, targets, and policy actions without averaging away critical tenants or turning SLOs into vanity dashboards.