Original practice scenario: Design a hypothetical manufacturing lineage service where a lot can split, merge or undergo rework. How would you answer which finished units share an upstream process step?
Read model answer +
Represent genealogy as immutable transformations between identified material units or lots. A single parent column cannot capture both splits and merges without losing provenance.
Work through it
Store each transformation with input identities, output identities, quantities, process-step version and event time; validate conservation rules appropriate to the process.
Build traversable upstream and downstream edges with indexes by material and transformation, retaining corrections as new versions rather than overwriting history.
Support bounded lineage queries and test partial splits, repeated rework and missing upstream events, returning incomplete evidence explicitly.
Tradeoffs to discuss
Fine-grained unit lineage improves precision but creates more records; lot-level lineage is cheaper while increasing the scope of affected material.
Common pitfalls
Prevent accidental cycles or distinguish a new rework generation from an impossible self-reference.
A complete graph does not establish that every source scan was recorded correctly; retain reconciliation and source-quality controls.
Original practice scenario: A hypothetical semiconductor analysis compares two process settings, but one setting has more inspected wafers and inspection selection depends on early warning signals. What would you examine?
Read model answer +
Inspection outcomes come from a selected population. Comparing observed failure percentages directly can confound process quality with the rule that decided which material to inspect.
Work through it
Define the experimental unit and eligible production population, then document inspection rates, selection triggers and outcome delays for each setting.
Prefer randomized or otherwise justified comparable assignment; stratify by product and process stage, and assess whether selection probabilities can be estimated credibly.
Report inspected yield and population coverage separately, with uncertainty and sensitivity analysis for unobserved outcomes rather than a single unsupported corrected estimate.
Tradeoffs to discuss
Additional random inspections improve identification but consume capacity; weighting can use existing data while depending on overlap and a defensible selection model.
Common pitfalls
Do not count multiple measurements of the same wafer as independent production units.
Check for product-mix changes and post-treatment inspection rules before attributing an apparent improvement to the process setting.
