Each question links back to the mental model behind it.
Start from the measured read path, freshness requirement, and failure budget. Then choose the cache location and invalidation strategy that make those constraints explicit.
A good key spreads writes, supports dominant access patterns, avoids hot tenants, and leaves a credible migration and rebalancing path.
Bound retries by an end-to-end deadline, use backoff and jitter, retry only safe or idempotent work, and shed load before retries amplify an outage.
Name the invariant and the user-visible cost of stale data. Pay coordination cost only where violating that invariant creates unacceptable harm.