MLSD-001 · DESIGN PRACTICE
Practice: Video recommendations
Draft the complete decision system before revealing the guidance. Use a 45-minute rehearsal budget and adapt it to the interviewer’s direction.
Preserve eligibility, available-at-decision inputs, mature labels, bounded work and compatible releases. Explain alternatives and test a normal, boundary and failure case.
Your practice draft
Design the home-page recommender under the stated 200 ms/2,000 req/s workload. Explain a complete baseline, label maturity and schema relationships, retrieval/ranking evaluation, compatible release and rollback. Then handle no-history consent, selective eligibility, a ranker outage and a same-dimension encoder change.
Saved locally on this device. The worked answer does not score your draft. Existing lesson notes and completion are separate.
Progressive hints
Hint 1
Start with unit, product objective, hard eligibility and a complete rule baseline; do not start with ANN brand names.
Hint 2
Draw decision → actual impression → interaction → label revision, then separate candidate recall from final rank quality.
Hint 3
Trace the full bundle and final live policy through the mismatch/outage; compute raw vector bytes and bounded pair work before claiming capacity.
Self-review evidence
For each criterion, record Missing, Prompted, Independent or Adaptive reasoning. No automated scoring or offer prediction is provided.
- Framing: decision/action/placement, assumptions and two changed-answer branches are concrete.
- Data: availability, consent, actual exposure, attribution, maturity/completeness, split and immutable lineage hold.
- Modeling: complete baseline, shape/objective/negative assumptions and a reason to keep the simpler design.
- Evaluation: distinct vector-overlap/relevance/ranking metrics with duplicate, tie and undefined-query semantics.
- Integration: explicit PK/FK/cardinality, compatible bundle, bounded work, eligibility and fallback.
- Operation: gated rollout/rollback, owners, delayed evidence, deletion limits and feasible cost assumptions.
Reveal worked design guidance
A coherent design uses an eligible followed/popular baseline, logs actual impressions separately from responses, builds only mature complete-source labels and uses available historical features. It introduces retrieval/ranking only with distinct metrics and a compatible artifact contract. The fixture computes overlap 2/3, relevance recall 1/3 and separate graded NDCG about .789596. It bounds scoring, rechecks eligibility, returns fewer items when necessary, rejects mixed bundles and retains a complete rollback. It estimates 5.12 GB raw item vectors and 200,000 expensive pairs/s at the declared caps without claiming observed capacity. Product lift and feedback correction require evidence beyond the toy fixture.
Open the complete worked design in a new tab →The new tab keeps this draft visible. Save it on this device before closing the page.