Original practice scenario: A hypothetical social application stores posts in an operational database, search index, caches and recommendation features. Design an inspectable deletion workflow.
Read model answer +
Enumerate the representations covered by the request and make deletion progress observable. Removing a primary row does not prove that copies, derived features or cached responses are gone.
Work through it
Authenticate the request, assign a durable deletion job identity and record affected resource identities without copying unnecessary personal content into logs.
Send idempotent deletion or invalidation commands to each covered store, retain completion receipts and retry failures with bounded backoff.
Prevent stale reingestion using deletion versions or tombstones, then reconcile outstanding stores and expose partial completion to the responsible operator.
Tradeoffs to discuss
Fast acknowledgment improves responsiveness but must distinguish request acceptance from completed removal; synchronous completion increases latency and dependency coupling.
Common pitfalls
Define how backups, legal retention and shared-content references are handled instead of silently promising universal immediate erasure.
A retrying data pipeline must not recreate content from an older snapshot after deletion.
Original practice scenario: A hypothetical feed-ranking model receives clicks, long views and hide actions. How would you incorporate negative feedback when a hide may refer to a topic, creator or presentation?
Read model answer +
Treat feedback as contextual evidence about a decision, not a universal label on a person or topic. Different actions have different ambiguity and observation biases.
Work through it
Record the shown item, candidate set, position and available context with the feedback, minimizing stored personal information and defining retention.
Distinguish explicit preference controls from inferred signals; compare interpretations using held-out behavior and independent quality judgments.
Evaluate user-level and content-slice outcomes, including repeated unwanted exposure and diversity, before a bounded online experiment with reversal controls.
Tradeoffs to discuss
Strong suppression respects a clear preference quickly but may narrow discovery; weaker updates adapt gradually while risking repeated frustration.
Common pitfalls
A missing click is not equivalent to a hide when the item may not have been noticed.
Do not evaluate only engagement gains; verify that the intended preference and user-control outcomes improve.
