Prevent cross-tenant side channels in a high-density shared service

• Define the adversary, observable signals, sensitivity of inferred activity, and contractual isolation tier. Separate cross-tenant data exposure from statistical workload inference and availability interference.

• Measure cache sharing, queue depth correlation, device scheduling, allocation granularity, and timing variance. Evaluate whether an attacker can repeat observations cheaply enough to extract useful signal.

• Prevent direct data mixing, reduce high-value inference below an accepted risk threshold, isolate abuse, and preserve shared utilization where safe. Stronger tenants need enforceable dedicated options.

• Use tenant-scoped cache keys, memory zeroing, fair queues, partitioned resource budgets, constant-shape errors, and coarse quota feedback. Place high-risk tenants in separate cells or hardware pools with independent telemetry.

• Tenant identity and isolation tier follow every cache, queue, allocation, and metric path. Aggregation delays or thresholds prevent exposing other tenants through real-time shared-capacity status.

• Isolation policy and placement are authoritative in the tenant directory; moves use epochs so no work continues in the old shared pool. Cache invalidation covers every tier after policy change.

• During overload, tenant-specific shedding avoids revealing a named neighbor and prevents one attacker from forcing global timing changes. Cleanup verifies device and memory state before reassignment.

• Apply defense in depth: authorization, namespace separation, resource isolation, output normalization, and penetration experiments. Do not claim constant time where shared systems cannot provide it.

• Internally track cross-tenant correlation, eviction and contention, placement violations, memory cleanup, and abuse probes. Customer telemetry reports only their allocation and service outcome.

• Partitioned caches and pools reduce hit rate and utilization; apply them by risk tier. Price dedicated hardware and quantify which mitigations deliver meaningful risk reduction.

• Random jitter can obscure simple timing but adds latency and often fails statistical attacks. Structural isolation is stronger; use noise only as a secondary measure.

• Fix direct namespace and memory issues first, introduce fair partitions and telemetry redaction, then offer cell placement for sensitive tiers. Continuously red-team representative shared workloads.

• Principal candidates avoid absolute claims and articulate threat, signal, mitigation strength, and cost. They distinguish logical tenancy from shared-resource leakage.

Interviewer evaluation rubric