Video-Keyword Association Review System

Design a service where users submit batches of 50+ videos and 1000+ keywords, then the system evaluates association between videos and keywords while supporting updates, versioning, and resumable upload. Requirements Design a backend module for checking associations between uploaded videos and keyword sets. The prompt is long and product-specific; the stable requirements are: A user can submit a batch containing 50+ videos and 1000+ keywords. The system evaluates the association / relevance between each video and the keyword set. Users can update a prior submission, so the design needs versioning rather than destructive overwrite. Uploads may be interrupted and resumed. The interviewer expects schema design and a typed design explanation even if no diagram tool is available. Key design decisions: Data model for submissions, submission versions, video assets, keyword sets, evaluation jobs, and per-version results. Whether to compute every video-keyword pair eagerly or use indexing / embedding retrieval to reduce candidate pairs before scoring. How resumable upload state is represented and how incomplete batches are isolated from active versions. How result recomputation works when only a subset of videos or keywords changes. Notes The prompt has both system-design and schema-design flavor. Start by normalizing entities and version ownership before drawing workers. The resource constraint is the cross product: 50 videos x 1000 keywords per submission, multiplied by repeated updates. A naive full recompute may be acceptable for small traffic but needs an incremental path for production scale. The interviewer may require a text-only design if the coding interface has no diagram canvas. Practice expressing tables, relationships, and dataflow in plain text. Preparation Drill versioned data models: immutable submission versions plus pointers to the active version. Prepare an ingestion pipeline with resumable upload, asset processing, association scoring, and result publication. Practice explaining eager all-pairs scoring vs. candidate-generation-plus-reranking trade-offs.