Skip to content
Tech Interview Prep
Technical interview guide

Data Mesh vs. Monolithic Warehouse

Centralized data ownership by one team versus federated, domain-owned data products, and the organizational tradeoff between them.

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

Scope: AWS Prescriptive Guidance and Analytics Lens, Google Cloud data mesh architecture series, Microsoft Purview governance-domain and data-product guidance, and Thoughtworks Data Mesh guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-4A team owns the `orders_curated` table, which 14 downstream teams read. Last quarter a column rename broke three of those pipelines, and the owning team found out nine days later from a Slack complaint. The team has budget for one change this quarter. Which change makes the dataset a data product?
MCQ-7A 30-domain logistics company runs a data mesh. Analysts spend days locating the canonical shipment-status table, two domains publish conflicting on_time_rate columns, and questions about a table's numbers go unanswered. Which practice is the anti-pattern responsible for these symptoms?
MCQ-24A retailer funded its data mesh to cut the time from a domain publishing a data product to a downstream team shipping a feature that uses it, with no rise in incident rates. Six months in, leadership asks which measure shows whether that goal was met. Which should the team report?
MCQ-18A 'Customer 360' data product is published by the Growth Analytics team. It joins three upstream data products: Orders (owned by Checkout, 1-hour freshness SLA), Tickets (owned by Support Platform, 15-minute SLA) and Billing (owned by Finance Data, 1-hour SLA). Customer 360 advertises a 2-hour freshness SLA and 99.9% availability. On Tuesday, Checkout ships an unannounced Orders schema change; Customer 360's load fails and its freshness breaches to 26 hours. Under data-product ownership conventions, who owns Customer 360's 2-hour SLA, and what does Checkout owe?
MCQ-15`orders.v3` is a data product with 14 downstream consumers. Its contract declares backward-compatible-only evolution within a major version: new optional fields are allowed, but no field may be removed, renamed, or have its type changed, and a 30-day deprecation window for any breaking change. A producer wants to rename `amount_cents` (integer) to `amount` and remove the old field entirely, because three internal dashboards have already switched. What does the contract require?
MCQ-12The Payments domain owns `payments.settlements`, consumed by Finance and Fraud. The published port exposes `txn_id`, `amount_minor`, `settled_at` and guarantees ≤ 15-minute freshness. The team's plan has three parts: (a) replace the batch pipeline with a streaming one, (b) keep the same three fields and the same 15-minute freshness, and (c) change `amount_minor` from cents to dollars because the new source stores dollars. Finance's ledger job multiplies `amount_minor` by 0.01. What should the team do?
MCQ-13Your team publishes `orders_clean` hourly. The fraud team's model reads it at 06:00 UTC and needs the newest order to be at most 4 hours old. Last month two silent DAG failures left the table 9 hours stale; the model scored stale data and flagged 3,100 legitimate orders. The platform dashboard still reports 99.98% pipeline success, and the team says the SLO is met. Which SLO change would have caught this failure?
MCQ-8In a data mesh, the `orders` data product is owned by the Orders domain team and declares freshness ≤ 15 minutes, a contractual commitment to an external partner. The domain's dbt job runs every 10 minutes and takes 3 minutes. The platform's Kafka ingestion for the orders topic shows a p99 lag of 22 minutes. The partner sees the table refresh every 25–30 minutes and escalates. The platform team is asked to fix the breach. What should it do?
MCQ-3A media company is carving data domains. It runs 9 source systems across 4 business capabilities, and the subscriber concept is written by 6 of those systems. A change to the entitlement rule currently takes five weeks to reach every system. The team proposes one domain per source system, each owning the data its system writes. Under that split, what happens to the subscriber concept?
MCQ-11A retailer has 14 domain teams. Its central data platform team owns one canonical physical model in the enterprise warehouse. Every new domain attribute requires a change request to that model; median lead time is 7 weeks, and 9 of 14 teams maintain private copies of customer and product tables to ship. Which diagnosis best explains why the canonical physical model is the bottleneck?
MCQ-2A payments company runs one central data platform: 40 engineers own ingestion, the warehouse and the BI layer for 9 domains. Leadership is weighing a move to domain-owned pipelines. Which fact about the current situation argues for keeping the centralized model?
MCQ-20A 400-engineer insurer is 18 months into a data mesh rollout across 11 domains. Each domain has a named data owner, but the central platform team still writes 80% of production pipelines and handles all pager alerts for data freshness. Which finding would confirm that domain ownership is nominal rather than real?
MCQ-10Northwind Retail runs 60 domain-owned datasets. Last quarter: a payments team renamed a column and 14 downstream consumers broke with no warning; an access audit found 9 of 12 recent leavers still had read access, because each domain ran its own ACL store; and one lineage question took 11 days because every domain described its tables differently. Leadership wants one change that fixes all three without freezing domain releases. Which change fixes all three failures while letting domains keep shipping independently?
MCQ-19A 40-domain mesh runs one central policy engine that masks PII at write time by matching column names against a rule list the platform team maintains. A domain's new pipeline wrote a column named `contact_addr`; the rule list only had `customer_email`, so unmasked addresses sat in the lake for 6 hours. The platform team has 4 engineers, and the mesh adds roughly 30 new columns a week. Which change prevents the next such miss?
MCQ-6A payments company runs a data mesh with 30 domain teams and roughly 400 registered data products. A global policy published as code in the shared policy engine requires every data product that exposes customer identifiers to tag them with the PII classification and mask them for all consumers. The fraud team's velocity model needs the raw identifiers to compute its features, and it asks the governance group to unblock it. Which change should the governance group make?
MCQ-1A retailer's central data platform team of 14 engineers supports 38 domain teams. Median lead time for a new domain dataset is 11 weeks, and 68% of the platform team's sprint capacity goes to domain-specific transformation requests. Compliance requires every PII-bearing dataset to enforce row-level access and to be deletable within 30 days of an erasure request; last quarter 3 of 9 erasure requests missed that window because lineage had to be traced by hand. Which change best addresses both the throughput bottleneck and the compliance failures?
MCQ-5A retail company runs a data mesh: 40 domain teams, and a central platform team of five engineers that provisions data products through a ticket queue. Median wait is 11 working days, and the five engineers spend about 80% of each sprint on that queue. An audit of the quarter's 240 tickets finds 187 of them are one of four shapes - batch ingest to the warehouse, streaming ingest from Kafka, warehouse transformation, and a BI extract. Hiring is frozen for the year. Fourteen domain teams have stopped filing tickets: they create the cloud resources themselves with hand-rolled IAM roles, and nine of their products, all built from those same four shapes, never reach the catalog. Which change ends the 11-day wait without adding to those nine unregistered products?
MCQ-16A platform team ships a self-service portal: golden-path templates for services, CI and deploys, plus a service catalog. Six months later, a time study of new-service onboarding shows a median 11 days to first production deploy. Of those 11 days, 8 are domain engineers writing IAM policies, registering lineage metadata by hand, and carrying on-call for the platform's own controllers. Which conclusion does the evidence support?
MCQ-9A retailer's central platform team of six built a golden-path pipeline covering build, test and deploy, then last quarter made it mandatory: any service not on the path needs a platform-team exception, and exceptions take about three weeks. Since then, distinct pipeline variants in the repo have risen from 6 to 19, and the platform team's feature-request queue is flat at roughly 140 open 'add support for X' items. What should the team expect next quarter if nothing changes?
MCQ-21A retailer's analytics pipeline resolves order lines to products by joining on `supplier_sku` alone against `product_xref(supplier_sku, internal_key, valid_from, valid_to)`, keeping the row with the greatest `valid_from` per SKU. Supplier SKU `A1234` was retired and later reused: it maps to `P-001` for 2023-01-01..2024-03-31 and to `P-778` from 2024-04-01 onward. A February 2024 order line for `A1234` is loaded on 2024-06-01. What does the pipeline do, and what follows?
MCQ-22A data mesh runs on one shared warehouse cluster that costs $120,000/month. The platform team allocates that cost to domains by query count. Domain A runs 40% of the queries but consumes 8% of cluster compute; Domain B runs 10% of the queries but consumes 55% of compute. After the first bill, Domain A's engineers start merging their many small metadata queries into fewer large ones. What should the platform team do?
MCQ-17`orders` is a 240-million-row data product with 14 downstream consumers: 9 dashboards, 3 reverse-ETL syncs, and 2 ML feature pipelines. `order_ts` is a string column and 4% of rows fail to parse in the nightly load, which is why the team is converting it to a timestamp. The plan: on day 0 the producer deploys code that dual-writes `order_ts` and a new nullable `order_ts_ts` for every incoming row; a batch backfill then converts the historical rows at 20 million rows per day; a row-level reconciliation then compares `order_ts` against `order_ts_ts` across all 240 million rows, takes 3 days, and is the gate every read path must wait for. The 9 dashboards and 3 reverse-ETL syncs can repoint their read paths as soon as the column is ready; the 2 ML feature pipelines cannot move before day 30. `order_ts` is scheduled to be dropped on day 60. The dashboards compare year over year across the full history. What is the earliest day the dashboards can repoint their read paths at `order_ts_ts`?
MCQ-23A retailer is moving 240 dashboards off a central warehouse onto a data mesh. 38 teams own the dashboards; 12 domain teams publish data products. Each domain must publish a contract and pass a nightly reconciliation against the warehouse before its consumers move. Two weeks in, the orders domain has published and 60 dashboards have moved, but its reconciliation shows a 0.4% row-count gap that appears only on days when the returns domain's feed lands after midnight. The lead wants to move the next 60 dashboards this weekend. What should the team do?
MCQ-14A retailer has 38 domains and a central data team of 9 with a 7-month backlog. It wants its first data mesh migration to prove domain ownership, cross-domain data contracts and federated governance before expanding. Which domain should go first?
MCQ-25A retailer's central data platform team of 14 engineers serves 60 domain teams. New-dataset requests wait a median of 7 weeks, and the shared warehouse accounts for 70% of the $410k monthly cloud bill. Finance needs one cost-allocation view and security needs one access-control model, but domain teams say only they understand their own metrics. Which target model fits these constraints?