Backfill a derived field while every source record keeps changing

• Define whether the field is a cache, evidence, or authoritative decision, and what freshness and false-positive rates are acceptable. Ask how algorithm versions are approved, explained, and rolled back.

• Assume billions of entities, a multi-week scan, high update churn, and limited source read budget. Partition by stable entity key and calculate whether recomputation can outrun live mutation for every cohort.

• Never present an unversioned value, avoid overwriting a result computed from newer source state, default safely when absent, and preserve the evidence used for each consequential decision.

• Write a versioned derivation record containing algorithm version, source version or watermark, computed time, confidence, and reason codes. Live change processing and batch backfill use the same deterministic computation service; a policy layer decides which versions may gate traffic.

• The backfill reads a source snapshot or version, computes, then conditionally writes only if the entity has not advanced beyond its input. Conflicts requeue from the latest source; live events continuously close the freshness gap.

• The derived value is valid only relative to a source version and algorithm version. The transaction path reads the value plus policy version and chooses allow, deny, or conservative review; it never treats missing as an accidental default.

• Checkpoint ranges, bound retries, quarantine deterministic failures, and monitor cohorts that cannot converge. A kill switch changes policy to the prior approved algorithm or review mode without deleting evidence.

• Restrict source features and derived explanations, prevent sensitive feature leakage in logs, and audit which version influenced each decision. Deletion must remove both derived values and retained feature evidence according to policy.

• Track coverage by tenant and cohort, source-version lag, conditional-write conflicts, computation failures, decision deltas between versions, and business harm metrics. Shadow the new version on live traffic before enforcement.

• Use incremental feature reads, priority cohorts, and shared batch/stream compute to avoid two inconsistent implementations. Retain full evidence only for regulated decisions; aggregate lower-risk diagnostics.

• Freezing source records would simplify proof but is impossible. Unconditional last-write-wins is cheap but can publish stale computation; source-version conditional writes create retries yet preserve truth.

• Deploy schema and readers first, then live derivation in shadow, then historical backfill, and only after coverage and quality gates enable enforcement by cohort. Keep old algorithm outputs through a defined rollback and appeal window.

• Distinguished candidates connect data protocol to policy, explainability, and business harm. They unify batch and streaming logic, version every input and decision, and design gradual enforcement rather than a binary launch.

Interviewer evaluation rubric