Data quality trade-offs

1. junior

   1. For an invalid row, compare deleting, fixing, flagging and quarantining.

   What it tests: Vocabulary of data quality actions.

   Model answer and scoring

   Deleting loses information and hides problems. Fixing is fine only with a rule that makes the correct value certain (like trimming padding). Flagging keeps the row in the data with a marker, good for legitimate but unusual values. Quarantining removes the row from clean data but keeps it, with a reason, for someone to fix.

   Strong answers

   • Clear criteria for each action.

   Red flags

   • Treats deletion as the default.

2. mid

   2. Why shouldn't amounts be stored as floating point, and what do you use instead?

   What it tests: Exact arithmetic for money.

   Model answer and scoring

   Binary floating point can't represent most decimal fractions exactly, so sums drift (0.1 + 0.2 isn't 0.3). For money I use a decimal type with a declared scale, such as DECIMAL(12,2) in SQL and decimal.Decimal in Python, or integer cents. Rounding happens only where the rules say, with a stated mode.

   Strong answers

   • Decimal or integer cents, explicit rounding.

   Red flags

   • Rounds floats at the end.

   Practise it: Sum amounts by currency, without converting

3. mid

   3. Amounts arrive as 1.234,50 and 1,234.50 in the same column. How do you parse them?

   What it tests: Locale-aware parsing.

   Model answer and scoring

   I need each row's locale: with de-DE the dot groups thousands and the comma is the decimal, with en-US it's the reverse. I validate each value against its locale's pattern (groups of exactly three digits) and quarantine rows with a missing or unknown locale instead of guessing.

   Strong answers

   • Per-row locale with validation.

   Red flags

   • Removes all punctuation.

   Practise it: Sum amounts by currency, without converting

4. senior

   4. Your pipeline works on the sample file. How do you gain confidence it will work on new data?

   What it tests: Testing beyond the visible data.

   Model answer and scoring

   I write the rules from the specification, not from the values I happen to see; test on hand-made edge cases (empty strings, missing keys, boundary values, ties); run on another batch; and check reconciliation and exception counts for surprises. A rule that only handles the visible values usually fails the next batch.

   Strong answers

   • Edge cases, new batches, reconciliation.

   Red flags

   • "It matched the sample."

5. junior

   5. When should you flag a suspicious value instead of removing it?

   What it tests: Preserving the population.

   Model answer and scoring

   Almost always when the value could be real, such as a very large sale. Flagging keeps the data and lets someone with context decide.

   I remove only values proven wrong by a rule, and keep them in an exception report.

   Strong answers

   • Defaults to flag.
   • Removes only rule failures.

   Red flags

   • Deletes outliers automatically.

   Practise it: Flag large transactions, keep refunds

6. mid

   6. Strict validation rejects 15% of rows. The business needs the data. What are your options?

   What it tests: Balancing strictness and usefulness.

   Model answer and scoring

   Keep the strict rule and report exceptions; relax a rule after the owner agrees it was too strict; or accept rows with a quality flag so users can filter.

   I quantify each option's effect on the key numbers and let the owner choose.

   Strong answers

   • Several options with impact.
   • Owner decides.

   Red flags

   • Silently relaxes rules.

   Practise it: Validate quantities and prices against ranges

7. senior

   7. Should a cleaning rule that fixes typos with fuzzy matching run automatically?

   What it tests: Automation risk.

   Model answer and scoring

   Only for low-risk fields with high-confidence matches, with a log of every change. For identities, money or categories that drive decisions, fuzzy changes go to review.

   I'd measure precision on a labelled sample before enabling it.

   Strong answers

   • Risk-based automation.
   • Measures precision; logs changes.

   Red flags

   • Turns it on everywhere.

   Practise it: Map channels with an approved dictionary

8. mid

   8. What data quality metrics would you track for a recurring feed?

   What it tests: Monitoring.

   Model answer and scoring

   Row counts per delivery, exception rate by rule, share of missing values in key fields, uniqueness of keys, and freshness.

   I'd alert on sudden changes, because a drift usually means a source change.

   Strong answers

   • Volume, validity, completeness, uniqueness, freshness.
   • Alerts on change.

   Red flags

   • Only checks that the file arrived.

9. junior

   9. Why should cleaning steps be reversible?

   What it tests: Reproducibility.

   Model answer and scoring

   Rules can be wrong. If raw data stays untouched and steps are code, I can fix a rule and rebuild instead of trying to undo edits.

   It also lets reviewers see exactly what changed.

   Strong answers

   • Raw stays immutable; steps as code.

   Red flags

   • Edits the source file directly.

10. senior

    10. How do you decide how much cleaning is enough?

    What it tests: Fitness for purpose.

    Model answer and scoring

    Enough is when the remaining issues don't change the decision the data supports. I test the sensitivity of the key numbers to the remaining problems.

    I document the known limitations instead of chasing perfection.

    Strong answers

    • Fitness for purpose.
    • Sensitivity check.

    Red flags

    • Aims for perfect data.

Model answers are original study material for practice. They are not employer-endorsed and don't guarantee an outcome.