Skip to content
Tech Interview Prep
Technical interview guide

Data Governance Fundamentals

Who owns which data, who can access it, and how quality and compliance are enforced across an organization.

Read
32 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed
Relevant for
Data Architect

Scope: UK Data Ownership Model 2026 and ESDA guidance, NIST Privacy Framework and CSF/RMF, EU GDPR, W3C DCAT 3, OpenLineage, and UK Data and AI Ethics Framework current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-3A governance council has signed off on the definition of 'active customer', the sensitivity classification of the customer table, and a seven-year retention rule. The work that remains is to enforce row-level permissions, encrypt the PII columns, and schedule the retention deletes. Which role owns that work?
MCQ-2A payments company's `transactions` table has a `merchant_category` column. The steward's quality rule flags rows carrying a retired 3-digit code; the latest scan shows 9% of rows. The domain owner's SLA for the column is 90% conformance. The steward maintains the definition and quality rules; the owner is accountable for the column's risk. What should the steward do?
MCQ-19A retailer's data governance council meets monthly, maintains a policy wiki, and reviews new data-sharing requests. In practice, engineers grant warehouse access through the platform's self-service UI without routing it to the council, and when a council decision conflicts with a product team's roadmap, the roadmap wins. Which anti-pattern does this describe?
MCQ-1`fct_orders` is owned by the Payments team; the Data Platform team operates the warehouse cluster and holds the write credentials. A currency-conversion bug has corrupted three years of `net_revenue`. The root cause is a missing `WHERE refunded = false` in the join, which double-counts refunds; the FX rate itself is correct. The fix rewrites 1.1B rows. Payments is at an offsite for a week; the CFO wants corrected numbers by Friday. The platform team can run the rewrite in 20 minutes. What should the platform team do?
MCQ-25A bank's audit sampled 200 datasets and found the same customer table labelled 'Confidential' by the retail unit and 'Internal' by the payments unit, with different access and retention applied in each. The bank wants a dataset's controls to follow from its label alone, wherever the dataset sits, and it must keep distinct obligations for personal, commercially sensitive and public data. Which approach should it take?
MCQ-4A retail bank's data governance council can certify one asset this week. It uses this risk matrix: Impact is High (3) if the asset feeds an external regulatory report or contains customer PII; Medium (2) if it feeds more than 50 internal consumers; Low (1) otherwise. Likelihood is High (3) if the asset had a quality incident in the last 30 days; Medium (2) if it had one in the last 90 days; Low (1) otherwise. Risk = Impact × Likelihood. Which asset should be certified first?
MCQ-17A bank scores every accepted risk as post-control likelihood x impact, each rated 1-5, and maps the product to the role that may sign: 1-6 the control owner, 7-12 the business unit director who owns the affected system and its remediation budget, 13-19 the CISO, 20-25 the board risk committee. Every sign-off expires after 90 days and cannot be delegated. A card-not-present fraud service keeps full card numbers in a legacy store reachable from the internet, and a public exploit is already in use, so analysts score exploitation likelihood 5 and impact 4. The planned tokenisation and network segmentation remove the reachable path, cutting exploitation likelihood to 2; they do not shrink the record set, so impact stays at 4. Who signs the acceptance of this residual risk?
MCQ-16An analytics warehouse tokenizes every table at ingest. The fraud team asks the governance board for an exception so its model can read raw account numbers, which it says costs 0.18 recall. The request states scope (the accounts and transactions tables), a reason, a compensating control (raw reads limited to the six-person fraud team, every query written to the audit log), an accountable owner of "the fraud team", a 90-day duration described as "renewed automatically if the model still needs it", and no convergence plan. Board rules: an exception is valid only if it states scope, reason, one named accountable owner, a compensating control that reduces the specific risk, and a plan for ending it; the term may not exceed 90 days and may not renew automatically. What should the board do?
MCQ-23A governance team reports to a steering committee. The dashboard shows 12,000 catalogued datasets, 11,400 with a named owner (95%), 3,100 of those owned datasets have had no owner activity in 12 months, 412 policies published this year against a 400 target, 98% steward training completion, 2,140 access requests processed, and a median 9 days to close critical data issues against a 3-day target. The committee asks for one number that shows whether ownership is operating, not merely assigned. Which figure belongs on the dashboard?
MCQ-22At a 900-person retailer, the central data governance team requires every new analyst dataset to pass a 12-day review by legal, security, and the data office before it can be queried in the warehouse. Analysts may not query the warehouse until approval. Last quarter, warehouse query volume from the merchandising analytics team fell 35%, and security found 47 unregistered CSV extracts on laptops and shared drives containing customer emails and order totals. The governance lead blames the 12-day review; the analytics director says analysts simply do not know the policy. The governance lead has one quarter to reduce unregistered extracts without dropping the review requirement. Which measurement would best distinguish between these two explanations?
MCQ-18A retailer's catalogue holds 1,200 entries, 180 tagged critical. An audit finds every critical entry has a non-null owner, but 96 of those owner values resolve to a shared mailbox decommissioned 14 months ago, and no critical entry has lineage. Last quarter, three incidents each took a median of 2.5 days to trace a broken downstream report back to its source table. The team has one sprint and plans to add a quality-score field to all 1,200 entries with a nightly completeness scan. Which action should the team take instead?
MCQ-6A warehouse's policy engine uses deny-overrides: if any policy matching a request says DENY, the request is denied; otherwise it is allowed if at least one matching policy says ALLOW; otherwise it is denied. The policy set for `warehouse.customers` is: - P1: ALLOW `role:analyst` SELECT when `env.network = corp-vpn` - P2: ALLOW `role:analyst` SELECT when `purpose = dashboard` - P3: DENY `role:contractor` SELECT - P4: DENY `*` SELECT when `env.network = public` A request arrives with roles `{analyst, contractor}`, resource `warehouse.customers`, action `SELECT`, `env.network = corp-vpn`, `purpose = dashboard`. What does the engine return, and why?
MCQ-8A payments platform's transactions table has 41 columns. The fraud team's queries filter only on merchant_id, amount, timestamp, and device_fingerprint; they never read cardholder_name, billing_address, email, or ip_address. Chargeback disputes need the full PAN in the clear for 13 months, and only the disputes team needs it; rows older than 13 months are already purged. The fraud team's service account holds SELECT on the whole table, and that grant cannot be narrowed this quarter. Which change should the team make?
MCQ-7A catalogue covers 9,000 datasets and serves search from an index of names, descriptions, column names and owners; the metadata is classified internal, approved for every employee. Reads are enforced at query time by the warehouse policy engine, which holds 1,400 grants. Six weeks ago the search service began filtering results through those same grants, so a dataset appears only if the searcher already holds a grant on it. Since then weekly distinct-dataset views per analyst have dropped from 38 to 7, 61% of searches return no results, and median time from search to first query has risen from 6 hours to 4 days. The grant count is unchanged. Analysts are still expected to request access to datasets they do not yet have. What should the team do?
MCQ-13A payments team keeps 4 TB of cardholder data in an S3 bucket encrypted with SSE-KMS, keys rotated every 90 days, all API traffic over TLS 1.3. A quarterly review finds that a support engineer used a standing read role to export 12,000 rows to a personal laptop, and that 60 contractors hold the same role. The CISO wants the finding closed because the data was encrypted at rest and in transit. Which remediation actually closes it?
MCQ-11A bank shares 24 months of card-transaction data with a vendor to train a fraud model. The agreement limits use to fraud detection, prohibits sharing the raw data with third parties, and requires deletion 90 days after the engagement. The vendor trains the model and licenses the model's API to a competing bank. The competing bank's fraud team can query the model and, in some cases, recover transaction details. Which action best addresses the leak?
MCQ-9A retailer replaces customer names with random tokens in its analytics warehouse. The mapping table sits in a separate, access-controlled service. Each row still carries full date of birth, full postcode and gender, and a published study finds those three fields uniquely identify 99.98% of the national population. A privacy engineer must classify the warehouse rows under GDPR. Which classification is correct?
MCQ-21A retailer collected 2.3 million customers' email addresses, order history and delivery addresses between 2019 and 2023 under a privacy notice that listed only 'processing and delivering your orders', relying on Art. 6(1)(b) contract necessity. The marketing team now wants to train a recommendation model on that same order history and email customers personalised offers. What does the GDPR require before that training can start?
MCQ-10A city plans to deploy a model that ranks 311 pothole repair requests by predicted severity. Legal confirms the open-data license permits this use. The ethics review board's charter says legal compliance is the floor, not the ceiling, and requires the team to examine harms, alternatives, and contestability before launch. Which analysis should the team add?
MCQ-24A warehouse table `orders_raw` holds order records dated 2016-01-01 through 2023-08-31 and stopped receiving writes on 2023-08-31. A legal retention policy requires every order record to be kept for 7 years from its order date. One downstream dashboard still queries the table; its team commits to migrating off by 2025-03-01. Today is 2024-06-30. Assuming no other obligations, what is the earliest date the entire table can be retired?
MCQ-14A service stores user events in a Postgres table partitioned by calendar month: events_2024_01 through events_2024_06. The retention policy is 90 days. A daily purge job enforces it by dropping whole partitions whose entire date range is older than the cutoff; it never deletes individual rows. The job has not run before. Today is 2024-06-15. Which partitions does today's run drop, and what is the oldest row still retained?
MCQ-12A team licenses a third-party sanctions-screening feed. The contract requires the vendor to issue corrections within 24 hours of a list change, and requires the recipient to apply each correction to every copy it controls within 5 days of receiving it. The team holds three copies: a raw landing zone updated in place, a warehouse table updated in place, and a snapshot rebuilt on the 1st of each month. A correction arrives on 26 March. What must the team do?
MCQ-5An `orders` table takes 200,000 rows a day and feeds two consumers. Finance reconciles revenue from `amount` and `currency` and needs the day's rows complete by 06:00. A recommender reads `order_id` and `user_id` and needs them within 5 minutes of the order. One nightly job runs 30 checks over the table and publishes a single row-weighted pass rate; last night it printed 99.98%. Finance still had to restate revenue because 12 rows had a null `currency` that the loader coerced to USD, and the recommender's copy was 40 minutes behind at peak. Which change makes the nightly number able to catch both failures?
MCQ-15A retailer builds column-level lineage by parsing the SQL in its dbt transformation repo. The graph shows the customer email column flowing into two internal dashboards. A separate Python job has been exporting the same column nightly to a marketing vendor's S3 bucket for 14 months; it assembles its SQL at runtime. A GDPR erasure request arrives for one customer. Which conclusion should the team act on?
MCQ-20`order_features` is rebuilt hourly from `orders` (retained 90 days) and holds a trailing 168-hour window of per-customer aggregates. A model scores hourly off that window. A deploy at Mon 00:00 double-counted shipping cost in `avg_order_value`; the bug ran until Wed 23:59 and was found and fixed seven days later, on the following Wed 23:59, when the 72 affected partitions were replayed. Which already-emitted model scores were computed on corrupted features and must be re-scored or flagged?