Homepage Viewport / Shelf Dedupe

Given Netflix homepage shelves, dedupe titles shown in the viewport. The prompt is deliberately vague: candidates must clarify horizontal shelf windows, vertical scrolling, global vs local dedupe, insertion in the middle, and production-scale trade-offs. Requirements Input: shelves: list[list[title_id]], where each inner list is one homepage shelf. Basic output: a transformed list of shelves with duplicate titles removed according to the agreed viewport rule. Common rule: while filling the first X visible slots of each shelf, avoid titles already visible globally; after visible slots are filled, dedupe only within the current shelf. Variant: horizontal and vertical scrolling both matter, so the visible viewport is not simply the first X items of every shelf. A frequently used concrete rule: dedupe titles within a row, but once a row already holds 6 distinct titles, later slots in that row may repeat. Follow-up: the per-row dedupe threshold is dynamic (could be 6 or any N), passed in as a parameter. Follow-up: only dedupe the first N rows; rows after that may repeat freely. Follow-up: conditional dedupe — certain special rows are exempt and skip dedupe entirely. Follow-up: a new shelf can be inserted in the middle; update the visible dedupe result. Follow-up: discuss production scale, cache locality, Redis / Bloom filter trade-offs, and false positives. Write test cases in the editor. Example shelves = [ ["A", "B", "C", "A"], ["B", "D", "E", "D"], ["C", "F", "A"] ] X = 2 # One common interpretation: # shelf 0 visible: A, B; tail dedup within shelf keeps C # shelf 1 visible skips B, keeps D, E; tail avoids local duplicate D # shelf 2 visible skips C and A, keeps F Notes The clean implementation uses two sets: global_visible across shelves and local_seen within the current shelf. The trap is adding a title to global_visible without also adding it to the local set. That allows duplicates later in the same shelf. Do not assume ranking changes are allowed. Several rounds explicitly ignore recommendation ranking and focus only on dedupe behavior. Pin down the replacement policy before coding — it is the most common clarification gap: which shelf keeps a contended title (top shelf usually wins by priority), and when a duplicate is removed from a visible slot, do you (a) shift the remaining titles forward, (b) pull a fresh title from the hidden tail of the shelf to backfill the slot, or (c) just leave a gap? Each choice changes the expected output. For production, a Bloom filter can reduce memory but introduces false positives. A false positive hides a valid title; be ready to decide whether that is acceptable. If middle insertion must be exact, global state after the insertion point may need recomputation. Incremental patching works only if the rule is local enough. Preparation Implement the global_visible + local_seen version cleanly. Create tests for duplicate within one shelf, duplicate across shelves, empty shelf, X = 0, and inserted middle shelf. Practice a 5-minute production discussion: cache scope, invalidation, recomputation cost, and approximate membership filters.