Skip to content
Tech Interview Prep
Technical interview guide

Master Data Management

Maintaining one authoritative version of core business entities (customers, products) across many systems.

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

Scope: IBM InfoSphere and watsonx.data MDM guidance, Microsoft Purview MDM guidance, W3C PROV-O, EU GDPR, and NIST Privacy Framework current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-1A retailer's MDM hub loads customer records nightly from three systems: e-commerce (12.4M rows), POS loyalty (9.1M rows), and a legacy CRM (3.3M rows). After the load the hub reports 24.8M member records and 14.6M entities. A downstream analyst counts rows in the entity table and reports 14.6M customers; the CRM team insists the company has 24.8M customers. Which explanation resolves the discrepancy?
MCQ-2A team merges customer rows from CRM, billing and support into a single entity. Auditors require that any surviving attribute value can be traced back to the exact source row and re-evaluated when that source changes. Which structure must the entity retain to satisfy this?
MCQ-3A customer hub ingests from two sources. CRM is priority 1, Billing is priority 2. Records merge when the normalized email matches exactly and name similarity scores at least 0.85. Survivorship takes each attribute from the highest-priority source, except phone, which always takes the value from the record with the latest updated_at. CRM holds: email alice.chen@acme.com, name "Alice Chen", phone +1-415-555-0101, updated_at 2024-03-01. Billing holds: email alice.chen@acme.com, name "Alicia Chen", phone +1-415-555-0199, updated_at 2024-05-14. The matcher scores the name pair 0.93. What does the hub publish as the golden record?
MCQ-9An MDM hub builds a golden record for a customer from three sources. The survivorship rule for the `phone` attribute is configured as: source priority CRM > Billing > Support, with recency used only to break ties between records from the same source. The current values are: CRM `+1-415-555-0101` (last updated 2021-03-01), Billing `+1-415-555-0199` (last updated 2024-06-15), Support `+1-415-555-0142` (last updated 2024-07-01). What does the golden record hold for `phone`?
MCQ-21A dedup job blocks 100,000 candidate record pairs and the matcher flags 10,000 of them as matches. An audit of 200 randomly chosen flagged pairs finds 180 true matches and 20 false merges; an audit of 200 randomly chosen non-flagged pairs finds 30 that are actually true matches. What is the matcher's estimated recall on this run?
MCQ-7A customer-dedup pipeline compares record pairs and merges any pair whose similarity score is at or above a threshold. A hand-labeled sample of 1,000 pairs breaks down as follows: | score band | pairs | true matches | | --- | --- | --- | | ≥ 0.90 | 80 | 78 | | 0.80 – 0.90 | 120 | 96 | | 0.70 – 0.80 | 200 | 15 | | 0.60 – 0.70 | 250 | 15 | | < 0.60 | 350 | 1 | A false merge costs about $25 in downstream cleanup; a missed match costs about $250 in duplicate outreach and support. Which threshold minimizes total cost on this sample?
MCQ-6A team deduplicates 8 million customer records. Their blocking key is `country + postal_code + first 4 characters of surname`; it yields 41 million candidate pairs, and against a hand-labelled sample the matcher scores 0.71 precision at 0.91 recall. To cut cost they extend the key to `country + postal_code + first 4 characters of surname + last 4 digits of phone`. Candidate pairs fall to 6 million, precision rises to 0.88, and recall falls to 0.63. The matcher's model, threshold and labelled sample are unchanged. What best explains the recall drop?
MCQ-5A dedupe job over 500,000 customer rows standardizes `address` in place (uppercase, strip punctuation, truncate to the first six characters) and the matcher scores that standardized field together with `zip5`. Before standardization the job blocked on the raw address and produced 307,500 blocks averaging 1.6 rows. After it, blocking on `standardized_address + zip5` produces 12,000 blocks averaging 41 rows, covering the 492,000 rows that have an address. Precision on the labeled sample falls from 0.96 to 0.71, and all 340 false merges are pairs whose standardized addresses are identical but whose full addresses differ after the sixth character: `1234 MAPLE ST` against `1234 MAPLE AVE`, `88 OAKRIDGE RD` against `88 OAKRIDGE LN`. The matcher's scoring and threshold are unchanged. Which statement accounts for the precision drop?
MCQ-8A dedup pipeline over 50M customer records uses a pairwise matcher that scores 0.97 precision and 0.95 recall on a uniformly sampled, double-annotated set of 20,000 pairs (annotator agreement 0.98). Blocking recall is 0.99 and the same threshold feeds connected-component clustering. The largest component holds 40,000 records that should be 12 customers. What best explains the gap?
MCQ-4A retailer merges a CRM export (one row per account; a company can hold several accounts) with a loyalty export (one row per person) to build a customer 360 table. The dedupe job reports a 12% duplicate rate, yet the merged table holds 3× as many 'customers' as the marketing team's list. What should the team do first?
MCQ-22Your identity graph resolves 40M profiles into 31M clusters. A proposed rule change drops the fuzzy-name threshold from 0.92 to 0.85. Replayed against a 2M-profile sample it adds 14,000 pairwise matches, and manual review of 500 of them finds 460 true duplicates. But the largest cluster in the sample grows from 9 profiles to 3,800, and 60% of the new matches land inside it. The downstream feature store caps a cluster's feature vector at 1,000 members. What should you do before shipping?
MCQ-10A customer master record merges phone numbers from three systems using latest-wins on each source's last_updated timestamp. CRM writes 09:00 UTC from an unverified web form. Billing writes 08:55 UTC after a successful phone challenge. Support writes 08:50 UTC from a manual ticket. All timestamps are UTC source update times; no clock skew is known. The golden record shows the CRM number. Which conclusion is justified?
MCQ-25A bank's MDM hub merges customer records from core banking, cards and online. It reports 4.2M source rows collapsed into 1.1M golden records and calls the program a success. An audit samples 500 golden records: 61 carry attribute values that no source system currently holds and no survivorship rule produces, and 9 pairs of golden records share a national ID but were never matched. Which reading of the audit is correct?
MCQ-16A master data hub builds golden customer records from CRM, billing, and support. To save 14 GB, the team drops the crosswalk table that maps each golden record to its contributing source keys (3 keys per record across 40 million records). Six weeks later, billing's nightly batch corrects 2,000 customer addresses. The hub's ingestion log shows 2,000 rows received and validated, 0 golden records updated, and no write events; the error log is empty; survivorship ranks billing above CRM for address. What explains the zero updates?
MCQ-11A support agent manually merges two duplicate customer profiles in an MDM system. The surviving golden record keeps alice@new.example; the losing record's alice@old.example is discarded. Fourteen months later, after two more merges on the same profile, an auditor asks who approved the original merge and what the email was before it. The team cannot answer because the tool retained only the golden record. Which configuration would have made that merge auditable without re-contacting the customer?
MCQ-12A retail MDM hub auto-merged two customer records into golden record GR-4471 because both carried the name 'Alice Chen' and the same phone number. The steward confirms they are different people: CRM-8812 (gold tier, 12 orders) and CRM-9903 (bronze tier, 2 orders). The loyalty service and the order service each keep a local copy of the golden record's tier and order count and update it only when the hub publishes a change event for that golden record. The loyalty service has already upgraded CRM-9903's owner to gold, and the order service shows 14 orders on both accounts. Which repair leaves the loyalty and order services showing the correct tier and order count for each customer?
MCQ-13A retailer runs four customer systems (CRM, ERP, billing and support) holding 12 million customer records with 3.4 million duplicates across them. The call centre needs one screen showing a customer's combined record, reflecting an edit made in any source within seconds. Each of the four business units must keep editing its own attributes, and none will accept another system overwriting its values. The funded budget covers a matching and cross-reference service, not a new golden-record store. Which design meets these constraints?
MCQ-14A bank's customer hub is meant to be the system of record for the 'risk rating' attribute. It stores only source record keys and a match score; every read queries the four source systems (core banking, cards, CRM, online) and picks a winning value at query time. An audit finds two different risk ratings returned for the same customer within the same second. Which mechanism explains the finding?
MCQ-15A retailer runs Salesforce for sales and SAP for finance, with a bidirectional integration that pushes all 4,200 account records to both systems every 15 minutes: 403,200 writes per system per day. Sales reps edit `Phone` in Salesforce; within 30 minutes it reverts to SAP's older value. Finance corrects `BillingAddress` in SAP; Salesforce's stale value overwrites it on the next run. Both teams must keep editing their own fields. Which change stops the reverts and the write churn?
MCQ-17A master-data service publishes one event per entity change to a Kafka topic keyed by source system; the consumer applies events through an 8-worker pool into a projection keyed on entity ID. After a bug fix the team replays the topic's 4.2M events from offset 0. The replay leaves 1,300 of 50,000 entities showing field values from earlier in the day, and a broker redelivery of 12 events increments a derived counter 12 extra times. Which change to the event contract and consumer removes both symptoms?
MCQ-18An order service publishes one Kafka event per order change. Each event carries a stable `event_id`, the `order_id`, a monotonically increasing `entity_version`, and `total_delta`, the change in that order's total since the previous event. The consumer, in one transaction, updates `orders` and runs `UPDATE customer_totals SET lifetime_value = lifetime_value + :total_delta`. The broker is at-least-once. A consumer-group rebalance redelivered order 8842's events v7 (+120.00), v8 (+30.00) and v9 (0.00, status-only) to a different consumer instance. Customer 51 owns orders 8842 (total 150.00) and 8850 (total 200.00), and its `lifetime_value` now reads 500.00 instead of 350.00. Which change makes the redelivered batch apply exactly once?
MCQ-19A telco builds an MDM hub that assigns one golden customer ID across three sources. Billing and CRM hold subscriber name, address, MSISDN and contract data collected for service delivery. A network-analytics store used for capacity planning holds session rows in which MSISDN and IMEI were replaced by surrogate keys; the analyst team's access was approved in writing on the basis that those rows are never joined to subscriber identity. The hub's matcher resolves 94% of the analytics rows to a golden ID that also carries billing identity. Which statement about the analytics data is correct?
MCQ-20An MDM hub runs entity resolution nightly. The matcher writes candidate pairs to a review queue; each queue row carries both source records' attribute payload as captured at 02:00 plus the match score. Stewards adjudicate the queue through a UI that authenticates as a shared service account with SELECT on the 90M-row source table, so a steward can search the whole table. Security has flagged the shared account and asked you to replace it with a per-steward grant. The blocking key is the first four characters of the normalized legal name plus the ISO country code, and 38% of blocking keys in the supplier domain also contain records owned by another business unit. Last quarter 1,200 merges were reversed because the reviewer approved a pair whose records had diverged since matching. Which grant do you implement?
MCQ-23An MDM program already governs Customer and Product master domains. Regulatory reporting needs one governed crosswalk between the internal country code, ISO 3166-1 alpha-2 and the legacy ERP country code, effective-dated so a code retired in 2019 still resolves correctly for 2018 transactions. The crosswalk is consumed by the reporting pipeline, the customer domain and the product domain. Which domain should own it?
MCQ-24A retailer is consolidating customer master data into a new hub fed by three sources: CRM, support and the loyalty system. The hub has run in shadow mode for 45 days: it computes a golden record nightly, but no consumer reads it yet. Reconciliation against the loyalty consumer, which reads name, email, phone and loyalty tier, reports: 2,311,000 of 2,400,000 records agree on all four attributes; 78,000 differ on email only, and in every one of those the hub's value came from the support system, which the loyalty consumer does not read and which was updated more recently; 11,000 differ on loyalty tier, because the hub's survivorship takes the value from the most recently updated mapped source and the loyalty system's tier field is not mapped into the hub yet. The team wants to point the loyalty consumer at the hub this week. What should it do?