B · Business understanding and domain modeling

Feature model

Feature definitions, time semantics, and reproducible computation.

Classification boundary

Include: Include feature definitions, event time, windows, backfills, and reproducible computation.

Choose a neighboring leaf when: Use data model for general entity schemas and C for business metric goals and causal conclusions.

This capability may appear in standards across different stages. Explore stage expectations

25 questions

Choose a group based on the evidence you need. Groups follow existing question formats, not difficulty or mandatory prerequisites. Each batch has up to eight questions.

Judgment and options25

Compare goals, constraints, options, and trade-offs.

Questions 1–8

An account may hold a base plan, temporary trial, campaign benefit, and support credit. Systems implement precedence differently, so access sometimes flips. How would you define eligibility and conflict rules?
A model feature will be retired, but several online models and historical reports still depend on it. How would you plan retirement and replacement?
An online feature must refresh every second, but it affects few requests and its maintenance cost keeps rising. How would you decide whether to keep it real time?
A risk model uses later default as its training label, but many transactions have no final outcome before the observation window ends. How would you handle label maturity, sample inclusion, and retrospective evaluation?
You are building a 'days active in the last 7 days' feature for a model that predicts next-day churn. How do you define its time semantics so training and online prediction mean the same thing?
Enterprise contracts specify first response, service restoration, and final resolution times, each with different pause conditions. Support tracks one SLA timer, causing disputes. How would you model this?
Recommendation and risk teams both need a “spend in the last 30 days” feature, but one uses calendar days and the other a rolling hour window, and refund handling differs. How would you decide whether the definition can be shared?
A risk value is usually present in training data, but live serving often substitutes a default and model performance fluctuates. How would you find whether the cause is data collection, timing, missing-value handling, or implementation differences?

Questions 9–16

Models, operational rules, and reports share a core feature, but the upstream business definition has changed and some consumers cannot migrate yet. How would you manage old and new definitions, ownership, and migration order?
An upstream dependency change will increase refresh latency for a costly feature set from seconds to ten minutes. How would you determine which decisions are affected, which features must remain real time, and when a fallback value is acceptable?
Two models call a feature with the same name but receive different values due to defaults and missing-data handling. How would you establish a contract and compatibility checks? Follow-up: If a critical feature refresh misses the real-time decision window, how would you choose among stale values, fallbacks, or refusing the decision?
A user clicks a campaign link on Monday and purchases on Tuesday, but the purchase event uploads on Friday because the client was offline. How would you define event time, arrival time, attribution windows, and backfill rules so results remain explainable?
A device reports a cumulative counter, but the counter resets after device replacement or data correction; the data platform may also backfill historical events. How would you define deltas, reset markers, negative differences, and recomputation boundaries?
A numeric feature value of zero may mean the user had no activity, missing may mean the device did not report, and unknown may mean computation is unfinished; the pipeline currently fills all three with zero. How would you define the semantics and downstream handling?
A global product counts daily activity by calendar day, but product analytics, the warehouse, and model serving use user time, UTC, and server time respectively; results often differ near month-end. How would you define boundaries and versioned semantics?
A user's transaction enters a behavioral feature window, then is reversed days later due to refund or fraud review; downstream models and reports have already consumed the feature. How would you model reversal, window recomputation, and impact propagation?

Questions 17–24

A team wants a ratio feature of completions divided by applications, but applications can be withdrawn, duplicated, or later invalidated beyond the observation window. How would you define numerator, denominator, maturity window, and recomputation semantics?
New users have little behavioral history, and the feature pipeline fills missing values with a global average, making “not observed” indistinguishable from “average behavior.” How would you represent cold-start status and evidence boundaries?
An online feature service can return its last computed value after a cache expires, but some predictions need the freshest feature while others tolerate brief staleness. How would you define freshness metadata, task constraints, and unavailable-value fallback?
A click-through feature suggests low interest in an item, but it appeared briefly at the bottom of a page; the team still trains a recommender on clicks. How would you model exposure opportunity, position, visibility duration, and user action separately?
A prediction formerly used individual history, but a product now recommends at the organization level; membership changes over time and a person may belong to several organizations. How would you define feature ownership, membership validity, and aggregation deduplication?
After an upstream migration, a feature keeps falling even though business activity is stable; consumers did not know the source field had been deprecated. How would you define feature lineage, health state, and replacement-source cutover?
After a feature definition is fixed, the modeling team wants a historical backfill. Training reproducibility requires the original version, while business reports want corrected values. How would you design both views?
Risk teams need to reproduce decision inputs from three months ago, but the online feature store keeps only the latest state to save cost. How would you balance reproducibility and storage cost?

Questions 25–25

A missing risk feature is automatically set to zero. Latency and error rates are stable, but denial rates changed significantly. How would you judge whether the default is safe?

Back to the whole domain