A retailer has 180 source systems, 14 business units, 6,000 reports, and a $12 million three-year modernization budget. How would you define the enterprise data target state and sequence capabilities?

I would define the target state as a capability map tied to measurable business journeys, then fund it in dependency order. - I would map inventory visibility, customer identity, finance close, and regulatory reporting to ingestion, governance, modeling, serving, and operations capabilities, naming one accountable owner for each. - The first 12 months would fund shared identity, metadata, CDC, and quality controls before migrating high-value reports, because those foundations remove repeated work across 14 units. - I would attach exit measures such as 95% lineage coverage, inventory freshness under 5 minutes, and retirement of 1,500 reports to quarterly investment gates.

A manufacturer stores 4 PB of sensor files, runs 20,000 SQL dashboards, and trains 80 models, with a 15-minute analytics SLA. Would you choose a warehouse, a lakehouse, or both?

I would use a lakehouse for sensor and ML workloads and a governed SQL serving tier for the dashboard workload. - Object storage with Iceberg or Delta keeps 4 PB economical and lets Spark engines process raw time-series data without copying it into warehouse storage first. - Curated dimensional tables would be published to Snowflake or a lakehouse SQL endpoint sized for 20,000 dashboards and isolated from model training compute. - I would keep one catalog, stable product contracts, and reconciliation checks across both tiers so the hybrid choice does not create two definitions of truth.

A financial company has 900 analysts, 12,000 governed SQL queries per hour, limited Spark skills, and a requirement to launch in six months. Would you select Snowflake or Databricks?

I would select Snowflake for this governed BI workload and six-month delivery constraint. - Its separated virtual warehouses let finance, risk, and ad hoc users isolate concurrency without asking a small platform team to tune Spark clusters. - Dynamic masking, row access policies, and mature SQL tooling fit the governed analyst population, while dbt can preserve reviewable transformations. - I would validate the choice with a two-week benchmark covering peak concurrency, a representative 20 TB schema, and projected annual credits rather than a feature checklist.

A media company ingests 25 TB daily, runs 300 Spark pipelines, trains recommendation models on GPUs, and needs streaming features within 60 seconds. Would you select Snowflake or Databricks?

I would select Databricks because Spark-native engineering, streaming, and ML dominate this workload. - Auto Loader and Structured Streaming can land 25 TB daily into Delta tables while sharing the same data with batch pipelines and feature generation. - Job clusters and GPU pools isolate 300 pipelines and model training, while Unity Catalog supplies central access and lineage controls. - I would benchmark end-to-end feature freshness, recovery from checkpoints, and DBU plus cloud cost, and keep BI serving separate if dashboard concurrency later dominates.

A company must expose 2 PB of lake tables to Spark, Trino, Flink, and two cloud vendors without vendor lock-in. Would you standardize on Delta Lake or Apache Iceberg?

I would standardize on Apache Iceberg because engine and cloud portability are explicit hard constraints. - Iceberg has broad native support across Spark, Trino, Flink, and managed cloud catalogs, so one table does not require a Databricks-centered execution path. - Hidden partitioning and partition evolution let teams change physical layout without breaking consumers as the 2 PB estate grows. - I would still prove row-level update throughput, catalog failover, and concurrent writer behavior with our engines before adopting one table format globally.

Thirty teams publish 600 tables through bronze, silver, and gold layers, and consumers require a 99.5% freshness SLO. What contracts would you put between the medallion layers?

I would make each layer an owned contract with different guarantees instead of treating medallion names as folder conventions. - Bronze preserves source payload, ingestion time, source offset, and replayability for 30 days, while accepting source schema drift into quarantined fields. - Silver guarantees typed keys, deduplication, PII classification, and declared lateness; gold guarantees business grain, metric definitions, and the 99.5% freshness SLO. - Promotion requires automated contract, reconciliation, and lineage checks, and a failed gold build keeps the last valid snapshot available to consumers.

An e-commerce platform processes 120,000 events per second, needs fraud features in 3 seconds and finance totals by 06:00, while retaining 7 years of data. Design the real-time plus batch platform.

I would use one durable event backbone with separate streaming and batch products sharing governed storage contracts. - Kafka would partition by account or order, and Flink would compute 3-second fraud features with checkpoints and idempotent sinks. - Raw events and CDC would land in Iceberg object storage for seven years, where Spark produces finance snapshots with source-to-ledger reconciliation before 06:00. - A shared schema registry, catalog, and business keys would connect both paths, while compute and failure domains remain isolated.

A logistics company has 40 event types, 15-minute late arrivals, a 5-second tracking SLA, and only eight platform engineers. Would you choose Lambda or Kappa architecture?

I would choose Kappa with replayable streams because one processing model fits the SLA and the eight-person team. - Kafka retains immutable events long enough for correction, and Flink event-time windows plus watermarks handle arrivals up to 15 minutes late. - Historical recomputation runs by replaying bounded Kafka or object-storage streams into a new versioned output instead of maintaining separate batch logic. - I would choose Lambda only if required backfills exceeded practical stream replay throughput or finance demanded a separately implemented batch control path.

You must capture changes from 220 PostgreSQL and Oracle databases at 70,000 rows per second with under 2-minute freshness and no source downtime. What CDC architecture would you use?

I would use log-based CDC into a partitioned event backbone, with source positions carried through every sink. - Debezium or native Oracle capture reads transaction logs without table scans, and connectors are grouped by source criticality so one noisy database cannot starve 219 others. - Kafka records source LSN or SCN, transaction identity, schema version, and tombstones; consumers deduplicate by source position and apply changes idempotently. - Initial snapshots use chunked consistent reads, then switch at a recorded log position, with count and checksum reconciliation before consumers adopt each feed.

A Kafka platform carries 300,000 events per second for 25 domains, and per-customer ordering must hold while no tenant may consume over 5% of capacity. How would you design topics and partitions?

I would partition domain-owned topics by a stable customer key and carry tenant identity separately for quotas and fairness. - Topic boundaries follow data products and retention classes, while partition counts come from measured broker and consumer throughput with 30% headroom. - The customer key preserves ordering only for events of that customer, not for the whole tenant; producer idempotence and a schema registry define delivery and compatibility. - Trusted tenant identity drives client quotas, separate consumer groups, and dedicated topics for the largest tenants, enforcing the 5% limit without changing customer ordering.

A payments stream runs at 45,000 events per second and must produce one ledger entry per transaction with a 4-second p99 SLA. What Kafka and Flink consistency contract would you design?

I would contract that each accepted Kafka transaction becomes one visible ledger row within 4 seconds p99 through an idempotent sink that publishes without waiting for a checkpoint. - Kafka producers use idempotence and transaction IDs, while Flink checkpoints offsets and keyed state every 10 to 30 seconds only for recovery. - The sink immediately upserts by transaction ID under a unique constraint, so replay after a checkpoint restores the same row rather than creating a second effect. - The contract states per-transaction uniqueness, ordering scope, 4-second p99 visibility, maximum recovery delay, and reconciliation from Kafka transaction IDs to ledger entries.

Spark compacts 80 TB of Iceberg tables nightly while Flink writes 20,000 updates per second, and readers require snapshot-consistent results. How would you coordinate the topology?

I would let both engines commit through the Iceberg catalog while separating write scope and resolving conflicts by snapshot retry. - Flink writes small equality deletes and data files to bounded partitions, while Spark compacts data and delete files only in partitions old enough to be outside the streaming lateness window. - The catalog provides atomic snapshot publication, so readers pin one snapshot instead of observing partially compacted files. - Maintenance uses commit retries, orphan-file cleanup after a safe retention period, and metrics for file count, conflict rate, and snapshot age.

A bank operates in the EU, US, and Singapore, processes 8 TB daily, and forbids raw customer data from leaving its legal region. Design the cross-region architecture.

I would keep raw and identified data inside regional planes and replicate only approved aggregates or tokenized records to a global plane. - Each region has its own ingestion, encryption keys, catalog policy boundary, and processing capacity, with residency tags enforced before any export. - Global products receive contract-approved columns with irreversible regional tokens or k-anonymous aggregates, not direct identifiers. - A central control plane distributes policy templates and metadata but stores no restricted payload, and residency tests block noncompliant pipelines before deployment.

An insurer must run analytics across AWS and Azure, move no more than 50 TB per month between clouds, and recover from one provider loss within 8 hours. What multi-cloud design would you choose?

I would use portable object-table formats with regional compute and replicate only recovery-critical products under the 50 TB cap. - Iceberg tables, open Parquet files, and a portable transformation layer avoid forcing routine reads across cloud boundaries. - Products are classified by recovery tier; only the datasets needed for the 8-hour objective receive incremental cross-cloud replication and catalog reconstruction manifests. - Quarterly failover tests rebuild identity, catalog, orchestration, and priority products in the second cloud, while egress budgets reject unapproved replication.

A marketplace needs sub-20 ms profile reads, ad hoc analysis over 3 PB, full-text search across 500 million listings, and an immutable audit trail. How would you choose storage technologies?

I would use polyglot storage because the four access patterns have incompatible latency and retention needs. - A key-value store serves profiles under 20 ms, Iceberg object tables hold 3 PB for analytical scans, and OpenSearch serves listing text queries. - A dedicated audit bucket uses WORM or Object Lock with a locked retention period, and signed hashes plus periodic integrity verification detect alteration; append-only writes alone are insufficient. - The architecture names one authoritative store per fact, defines rebuild paths for search and caches, and prevents dual writes through events or an outbox.

A global distributor must integrate 140 volatile sources, preserve history for 10 years, and onboard acquisitions within 30 days. How would you design a Data Vault 2.0 core?

I would use a Data Vault 2.0 raw core to absorb source change quickly while preserving auditable history. - Hubs represent stable business keys such as customer and product, links capture relationships, and satellites keep source-specific descriptive history with load time and record source. - Hash keys allow parallel loading across 140 sources, while business-key collision rules and same-as links are governed centrally rather than hidden in loaders. - Acquisitions first map keys and source metadata into the raw vault within 30 days; business rules remain in a business vault and downstream marts.

The Data Vault core contains 12 years of sales history, but 1,200 analysts need daily revenue queries under 5 seconds. How would you build dimensional marts without creating a second truth?

I would publish star-schema marts as versioned projections of the vault, not as independently mastered data. - Business Vault point-in-time and bridge tables resolve history efficiently, then marts expose a declared order-line grain, conformed dimensions, and SCD behavior. - Incremental builds use vault load timestamps and reconcile revenue, row counts, and key coverage back to the raw core on every release. - Metric definitions and mart contracts live in the semantic catalog, while the 12-year vault remains the auditable source for reconstruction.

A company has 18 product teams spanning commerce, fulfillment, finance, and customer support, with shared customer and order data. How would you draw data mesh domain boundaries?

I would draw boundaries around business capabilities and decision rights, not existing application databases. - Commerce owns offer and order intent, fulfillment owns shipment execution, finance owns posted monetary events, and customer owns mastered customer identity attributes. - Shared identifiers and lifecycle events cross boundaries through versioned products, while each product states grain, authority, freshness, and allowed uses. - I would test the map against five real analytical journeys and split a domain only when one owner cannot meet distinct change cadence or policy obligations.

Eight domains publish 240 data products to 600 consumers, and breaking changes require 90 days notice. What would a platform-enforced data product contract contain?

I would make the contract machine-readable and enforce identity, semantics, service levels, policy, and compatibility at publication time. - It declares owner, grain, keys, schema, classifications, freshness, availability, retention, and support channel for each of the 240 products. - A registry checks backward compatibility, quality thresholds, lineage, and 90-day deprecation metadata in CI before a producer can publish. - Consumer telemetry identifies affected users, and incompatible change requires a parallel major version with migration evidence rather than an in-place edit.

Twelve domains operate independently, but regulators require one PII classification policy and audit evidence within 48 hours. How would you implement federated governance?

I would centralize mandatory policy and evidence standards while delegating product decisions and remediation to domain owners. - A central council defines PII taxonomy, retention classes, control templates, and the 48-hour evidence format, with legal and security holding veto rights. - Domain stewards classify assets, approve access, and own SLOs through policy-as-code embedded in their delivery pipelines. - The platform aggregates control results, lineage, exceptions, and ownership into one audit view, and time-bound exceptions require a named risk owner.

Data Architect interview questions

100 real questions with model answers and explanations for Senior Data Architect candidates.

See a Data Architect resume example →

Practice with flashcards

Spaced repetition · Hunter Pass

Questions

system-design

I would define the target state as a capability map tied to measurable business journeys, then fund it in dependency order.

I would map inventory visibility, customer identity, finance close, and regulatory reporting to ingestion, governance, modeling, serving, and operations capabilities, naming one accountable owner for each.
The first 12 months would fund shared identity, metadata, CDC, and quality controls before migrating high-value reports, because those foundations remove repeated work across 14 units.
I would attach exit measures such as 95% lineage coverage, inventory freshness under 5 minutes, and retirement of 1,500 reports to quarterly investment gates.

Why interviewers ask this: The interviewer is testing whether the candidate converts an enterprise vision into owned capabilities, dependencies, funding stages, and measurable outcomes.

sqlwarehouselakehouse

I would use a lakehouse for sensor and ML workloads and a governed SQL serving tier for the dashboard workload.

Object storage with Iceberg or Delta keeps 4 PB economical and lets Spark engines process raw time-series data without copying it into warehouse storage first.
Curated dimensional tables would be published to Snowflake or a lakehouse SQL endpoint sized for 20,000 dashboards and isolated from model training compute.
I would keep one catalog, stable product contracts, and reconciliation checks across both tiers so the hybrid choice does not create two definitions of truth.

Why interviewers ask this: The interviewer is evaluating whether the candidate chooses architectural boundaries from workload economics and service guarantees rather than labels.

sqlqueriessnowflake

I would select Snowflake for this governed BI workload and six-month delivery constraint.

Its separated virtual warehouses let finance, risk, and ad hoc users isolate concurrency without asking a small platform team to tune Spark clusters.
Dynamic masking, row access policies, and mature SQL tooling fit the governed analyst population, while dbt can preserve reviewable transformations.
I would validate the choice with a two-week benchmark covering peak concurrency, a representative 20 TB schema, and projected annual credits rather than a feature checklist.

Why interviewers ask this: The interviewer is checking whether the candidate makes a platform choice from workload, skills, delivery time, governance, and measured cost.

snowflakestreamingci-cd

I would select Databricks because Spark-native engineering, streaming, and ML dominate this workload.

Auto Loader and Structured Streaming can land 25 TB daily into Delta tables while sharing the same data with batch pipelines and feature generation.
Job clusters and GPU pools isolate 300 pipelines and model training, while Unity Catalog supplies central access and lineage controls.
I would benchmark end-to-end feature freshness, recovery from checkpoints, and DBU plus cloud cost, and keep BI serving separate if dashboard concurrency later dominates.

Why interviewers ask this: The interviewer is testing whether the candidate can distinguish a compute-intensive ML platform decision from a conventional BI procurement choice.

procurement

I would standardize on Apache Iceberg because engine and cloud portability are explicit hard constraints.

Iceberg has broad native support across Spark, Trino, Flink, and managed cloud catalogs, so one table does not require a Databricks-centered execution path.
Hidden partitioning and partition evolution let teams change physical layout without breaking consumers as the 2 PB estate grows.
I would still prove row-level update throughput, catalog failover, and concurrent writer behavior with our engines before adopting one table format globally.

Why interviewers ask this: The interviewer is evaluating whether the candidate ties an open table format choice to interoperability, evolution, and tested concurrency behavior.

slo

I would make each layer an owned contract with different guarantees instead of treating medallion names as folder conventions.

Bronze preserves source payload, ingestion time, source offset, and replayability for 30 days, while accepting source schema drift into quarantined fields.
Silver guarantees typed keys, deduplication, PII classification, and declared lateness; gold guarantees business grain, metric definitions, and the 99.5% freshness SLO.
Promotion requires automated contract, reconciliation, and lineage checks, and a failed gold build keeps the last valid snapshot available to consumers.

Why interviewers ask this: The interviewer is checking whether the candidate defines enforceable semantics and reliability at layer boundaries rather than merely naming layers.

designbatchconcurrency

I would use one durable event backbone with separate streaming and batch products sharing governed storage contracts.

Kafka would partition by account or order, and Flink would compute 3-second fraud features with checkpoints and idempotent sinks.
Raw events and CDC would land in Iceberg object storage for seven years, where Spark produces finance snapshots with source-to-ledger reconciliation before 06:00.
A shared schema registry, catalog, and business keys would connect both paths, while compute and failure domains remain isolated.

Why interviewers ask this: The interviewer is evaluating an end-to-end design that meets two latency classes without duplicating truth or coupling failure domains.

lambda

I would choose Kappa with replayable streams because one processing model fits the SLA and the eight-person team.

Kafka retains immutable events long enough for correction, and Flink event-time windows plus watermarks handle arrivals up to 15 minutes late.
Historical recomputation runs by replaying bounded Kafka or object-storage streams into a new versioned output instead of maintaining separate batch logic.
I would choose Lambda only if required backfills exceeded practical stream replay throughput or finance demanded a separately implemented batch control path.

Why interviewers ask this: The interviewer is testing whether the candidate weighs operational duplication against replay limits and team capacity.

databasepostgresarchitecture

I would use log-based CDC into a partitioned event backbone, with source positions carried through every sink.

Debezium or native Oracle capture reads transaction logs without table scans, and connectors are grouped by source criticality so one noisy database cannot starve 219 others.
Kafka records source LSN or SCN, transaction identity, schema version, and tombstones; consumers deduplicate by source position and apply changes idempotently.
Initial snapshots use chunked consistent reads, then switch at a recorded log position, with count and checksum reconciliation before consumers adopt each feed.

Why interviewers ask this: The interviewer is evaluating source safety, ordering, replay, snapshot cutover, and end-to-end correctness at enterprise CDC scale.

kafkadesigncapacity

I would partition domain-owned topics by a stable customer key and carry tenant identity separately for quotas and fairness.

Topic boundaries follow data products and retention classes, while partition counts come from measured broker and consumer throughput with 30% headroom.
The customer key preserves ordering only for events of that customer, not for the whole tenant; producer idempotence and a schema registry define delivery and compatibility.
Trusted tenant identity drives client quotas, separate consumer groups, and dedicated topics for the largest tenants, enforcing the 5% limit without changing customer ordering.

Why interviewers ask this: The interviewer is checking whether the candidate balances domain ownership, ordering, parallelism, and noisy-neighbor control in Kafka topology.

kafkadesignconsistency

I would contract that each accepted Kafka transaction becomes one visible ledger row within 4 seconds p99 through an idempotent sink that publishes without waiting for a checkpoint.

Kafka producers use idempotence and transaction IDs, while Flink checkpoints offsets and keyed state every 10 to 30 seconds only for recovery.
The sink immediately upserts by transaction ID under a unique constraint, so replay after a checkpoint restores the same row rather than creating a second effect.
The contract states per-transaction uniqueness, ordering scope, 4-second p99 visibility, maximum recovery delay, and reconciliation from Kafka transaction IDs to ledger entries.

Why interviewers ask this: The interviewer is evaluating whether the candidate distinguishes processing claims from externally observable correctness guarantees.

snapshot

I would let both engines commit through the Iceberg catalog while separating write scope and resolving conflicts by snapshot retry.

Flink writes small equality deletes and data files to bounded partitions, while Spark compacts data and delete files only in partitions old enough to be outside the streaming lateness window.
The catalog provides atomic snapshot publication, so readers pin one snapshot instead of observing partially compacted files.
Maintenance uses commit retries, orphan-file cleanup after a safe retention period, and metrics for file count, conflict rate, and snapshot age.

Why interviewers ask this: The interviewer is checking whether the candidate understands table-format concurrency, snapshot isolation, and the interaction between streaming writes and compaction.

designconcurrency

I would keep raw and identified data inside regional planes and replicate only approved aggregates or tokenized records to a global plane.

Each region has its own ingestion, encryption keys, catalog policy boundary, and processing capacity, with residency tags enforced before any export.
Global products receive contract-approved columns with irreversible regional tokens or k-anonymous aggregates, not direct identifiers.
A central control plane distributes policy templates and metadata but stores no restricted payload, and residency tests block noncompliant pipelines before deployment.

Why interviewers ask this: The interviewer is evaluating whether the candidate separates control and data planes while making residency enforcement architectural rather than procedural.

design

I would use portable object-table formats with regional compute and replicate only recovery-critical products under the 50 TB cap.

Iceberg tables, open Parquet files, and a portable transformation layer avoid forcing routine reads across cloud boundaries.
Products are classified by recovery tier; only the datasets needed for the 8-hour objective receive incremental cross-cloud replication and catalog reconstruction manifests.
Quarterly failover tests rebuild identity, catalog, orchestration, and priority products in the second cloud, while egress budgets reject unapproved replication.

Why interviewers ask this: The interviewer is testing whether the candidate treats multi-cloud as selective resilience with quantified data movement rather than full duplication.

searchimmutability

I would use polyglot storage because the four access patterns have incompatible latency and retention needs.

A key-value store serves profiles under 20 ms, Iceberg object tables hold 3 PB for analytical scans, and OpenSearch serves listing text queries.
A dedicated audit bucket uses WORM or Object Lock with a locked retention period, and signed hashes plus periodic integrity verification detect alteration; append-only writes alone are insufficient.
The architecture names one authoritative store per fact, defines rebuild paths for search and caches, and prevents dual writes through events or an outbox.

Why interviewers ask this: The interviewer is evaluating whether the candidate selects stores by access pattern while preserving authority and rebuildability across systems.

designdata-vaultonboarding

I would use a Data Vault 2.0 raw core to absorb source change quickly while preserving auditable history.

Hubs represent stable business keys such as customer and product, links capture relationships, and satellites keep source-specific descriptive history with load time and record source.
Hash keys allow parallel loading across 140 sources, while business-key collision rules and same-as links are governed centrally rather than hidden in loaders.
Acquisitions first map keys and source metadata into the raw vault within 30 days; business rules remain in a business vault and downstream marts.

Why interviewers ask this: The interviewer is checking whether the candidate uses Data Vault structures to solve volatility, parallel ingestion, auditability, and rapid onboarding.

queriessecretsdata-vault

I would publish star-schema marts as versioned projections of the vault, not as independently mastered data.

Business Vault point-in-time and bridge tables resolve history efficiently, then marts expose a declared order-line grain, conformed dimensions, and SCD behavior.
Incremental builds use vault load timestamps and reconcile revenue, row counts, and key coverage back to the raw core on every release.
Metric definitions and mart contracts live in the semantic catalog, while the 12-year vault remains the auditable source for reconstruction.

Why interviewers ask this: The interviewer is evaluating whether the candidate combines an auditable integration core with performant dimensional consumption and reconciliation.

data-mesh

I would draw boundaries around business capabilities and decision rights, not existing application databases.

Commerce owns offer and order intent, fulfillment owns shipment execution, finance owns posted monetary events, and customer owns mastered customer identity attributes.
Shared identifiers and lifecycle events cross boundaries through versioned products, while each product states grain, authority, freshness, and allowed uses.
I would test the map against five real analytical journeys and split a domain only when one owner cannot meet distinct change cadence or policy obligations.

Why interviewers ask this: The interviewer is checking whether the candidate can establish domain authority while handling shared entities without centralizing every model.

versioning

I would make the contract machine-readable and enforce identity, semantics, service levels, policy, and compatibility at publication time.

It declares owner, grain, keys, schema, classifications, freshness, availability, retention, and support channel for each of the 240 products.
A registry checks backward compatibility, quality thresholds, lineage, and 90-day deprecation metadata in CI before a producer can publish.
Consumer telemetry identifies affected users, and incompatible change requires a parallel major version with migration evidence rather than an in-place edit.

Why interviewers ask this: The interviewer is evaluating whether the candidate turns data product language into enforceable producer and consumer obligations.

classificationpiidata-governance

I would centralize mandatory policy and evidence standards while delegating product decisions and remediation to domain owners.

A central council defines PII taxonomy, retention classes, control templates, and the 48-hour evidence format, with legal and security holding veto rights.
Domain stewards classify assets, approve access, and own SLOs through policy-as-code embedded in their delivery pipelines.
The platform aggregates control results, lineage, exceptions, and ownership into one audit view, and time-bound exceptions require a named risk owner.

Why interviewers ask this: The interviewer is checking whether the candidate separates nonnegotiable enterprise controls from domain-level accountability and execution.

Locked questions

21
An enterprise has 2,500 datasets across five platforms and wants automated discovery in nine months, but business ownership must remain in 20 domains. Is this data fabric, data mesh, or both?
ownershipdata-mesh
22
A healthcare group has 65 million patient records, 14 source systems, and a requirement to create a golden record with fewer than 0.1% false merges. How would you design MDM?
system-designdesign
23
Order status differs between an operational system, the warehouse, and the MDM hub for 3% of 40 million orders. Which system should define truth under a 10-minute reporting SLA?
warehousesystem-design
24
A company has 7,000 tables, 1,500 users, and three catalog products under consideration, with a mandate to reach 90% ownership coverage in six months. How would you select and roll out a metadata catalog?
metadata-managementdata-catalogcoverage
25
Regulators require column-level lineage from 35 source systems through 900 transformations to 120 reports, with evidence produced within 24 hours. How would you architect it?
system-designlineageschema
26
Finance and product teams have six definitions of active customer across 400 dashboards, and metric queries must return within 3 seconds. How would you design a semantic metrics layer?
queriesdesignmonitoring
27
Ten teams operate 320 critical datasets, and executives require 99.7% on-time delivery plus detection of material quality failures within 10 minutes. Design the observability and SLO model.
designsloobservability
28
Twenty domain teams publish customer products, but the platform team has only six engineers and cannot author every quality rule. How would you divide quality responsibilities while maintaining a 99.5% SLO?
slo
29
A customer invokes GDPR erasure across 60 systems, 1.2 PB of lake data, and immutable financial records that must remain for seven years. Design deletion handling with a 30-day deadline.
system-designdesignimmutability
30
A telecom estate grows by 90 TB monthly across raw events, curated tables, logs, and backups, with retention rules of 30 days, 2 years, and 7 years. How would you architect retention?
retentionbackups
31
A global HR warehouse serves 30 countries and 4,000 managers, who may see only their hierarchy and cannot view salary columns outside compensation teams. How would you design IAM?
schemawarehousedesign
32
Five analytics platforms need to join customer behavior without exposing email or phone, and security requires token revocation within one hour. How would you design tokenization?
design-systemjoinsdesign
33
Forty producers feed 600 consumers, schemas change weekly, and consumers require 90 days to migrate. What schema evolution policy would you enforce?
schema
34
A bank publishes 75 event types to 160 applications, and event history must remain replayable for seven years despite contract changes. How would you evolve event contracts?
35
Airflow must schedule 6,000 daily jobs for 25 teams, with finance finishing by 06:00 and research allowed to consume only 20% of compute. How would you design orchestration and isolation?
designorchestrationairflow
36
A shared platform serves 700 BI users, 120 ELT jobs, and 40 ML teams, with dashboard p95 under 4 seconds and a fixed monthly compute budget. How would you isolate workloads?
etlelt
37
A claims platform stores 600 TB, requires RPO 15 minutes and RTO 2 hours, and must survive loss of an entire region. What disaster recovery architecture would you approve?
architecture
38
An on-prem Oracle warehouse holds 450 TB, supports 3,200 reports, and must move to cloud in 18 months with no reporting outage over 30 minutes. How would you stage the migration?
migrationswarehouse
39
A mainframe emits 8 million fixed-width records nightly, has a four-hour batch window, and contains 30 years of COBOL business rules that cannot all be rewritten in year one. How would you modernize its data architecture?
batch
40
Two merged companies have 9 PB of data, overlapping customer IDs, 11 BI tools, and a board deadline of 12 months for consolidated revenue reporting. How would you merge the estates?
estimation
41
A SaaS analytics product serves 12,000 tenants from 20 GB to 80 TB each, requires p95 under 6 seconds, and forbids cross-tenant exposure. How would you design isolation?
designcloud
42
Commerce, inventory, and finance domains update one order asynchronously, yet daily revenue must reconcile to within $10 across 50 million orders. How would you design cross-domain consistency?
designconsistencyasync
43
Forty ML teams need online features under 25 ms, offline training over two years, and point-in-time correctness, with 5 TB of new feature data daily. How would you architect feature data?
44
A new platform expects growth from 60 TB to 1.5 PB and from 2,000 to 80,000 events per second in 24 months, while procurement needs a capacity plan now. How would you size it?
procurementcapacity
45
Cloud data spend is capped at $350,000 per month for 30 teams, including 4 PB storage and 18 million query minutes. How would you design FinOps controls without blocking delivery?
queriesdesignfinops
46
Twenty teams need dev, test, and prod data-platform environments in two regions, and regulated resources require reproducible approval evidence. How would you structure Terraform?
reproducibilityterraform
47
Six teams disagree on Delta versus Iceberg for a 3 PB multi-engine lake, and a decision is required in three weeks. How would you run the RFC and ADR process?
conflictdecision-makingconcurrency
48
A new contract platform is technically ready, but only 3 of 16 domains use it and migration must reach 80% within two quarters. What architecture adoption plan would you lead?
decision-makingmigrationsarchitecture
49
During review of a proposed 500 TB customer lake, three mid-level architects each chose a different partition and identity strategy, and design approval is due in five days. How would you mentor through the review?
mentoringdesignpartitioning
50
A customer-360 product must serve 1,000 analysts and 35 operational APIs, with analytical queries under 5 seconds and API reads under 100 ms over 2 billion profiles. How would you design the serving architecture?
designapiqueries
51
Your lakehouse bill jumps from $180,000 to $410,000 in one month while query volume rises only 12%; what do you do in the first day?
querieslakehouse
52
At 09:00, Snowflake dashboard latency rises from 8 seconds to 95 seconds and credits burn at triple the normal rate across 300 analysts; how do you respond?
snowflakelatency
53
A new BigQuery dashboard scans 1.6 PB per day and adds $8,000 to the daily bill after its first release; what evidence and controls do you use?
bigquery
54
A 12-node Redshift cluster reaches 100% disk on two nodes while the others remain below 45%, and the nightly load misses its 06:00 SLA; how do you handle it?
soft-skillsredshift
55
Kafka consumer lag on an orders topic grows from 20,000 to 14 million records in 25 minutes, but producer throughput is unchanged at 90,000 events per second; what is your incident plan?
zero-to-onekafkathroughput
56
One of 120 Kafka partitions receives 38% of all payment events after a large merchant launch, causing 11-minute latency; how do you correct the hot partition without losing ordering?
partitioningkafkalatency
57
A Kafka consumer group handling 240 partitions rebalances every 3 minutes after a deployment, and processing availability falls to 62%; what do you inspect and change?
partitioningkafkadeployment
58
After a database failover, CDC has a 47-minute gap covering about 6.2 million customer updates; how do you recover without overwriting newer warehouse state?
databasewarehouse
59
A connector restart duplicates 18 million CDC events and inflates daily revenue by 7%; how do you contain and repair the incident?
incidents
60
A CDC pipeline has ignored deletes for 12 days, leaving 3.4 million supposedly removed profiles in analytics; what is your response?
ci-cd
61
Debezium replication slots retain 4.8 TB of PostgreSQL WAL and production disk will fill in 3 hours; what do you do?
postgresreplication
62
A Flink watermark change from 30 minutes to 5 minutes drops 2.1% of mobile events in regions with weak connectivity; how do you restore correctness?
63
A Spark join processes 9 TB but one task runs for 74 minutes while the median task finishes in 80 seconds; how do you diagnose and fix it?
joinsconcurrency
64
A Spark pipeline starts losing executors to OOM after input grows from 3 TB to 5 TB, and three retries have already consumed 900 core-hours; what is your next move?
ci-cd
65
Planning an Iceberg query now takes 42 seconds because a table has 310,000 snapshots and 18 million manifest entries; how do you recover performance safely?
queriessnapshotperformance
66
An Iceberg table receives 40 million 20 KB files per day from streaming jobs, pushing query p95 from 12 seconds to 4 minutes; what do you change?
queriesstreaming
67
Two Delta Lake jobs concurrently update the same 60 partitions, producing 1,900 commit conflicts in an hour and delaying customer balances; how do you resolve it?
partitioningconcurrency
68
A table-format upgrade applied to 900 TB of lake data makes 14% of legacy readers fail schema resolution; how do you manage the incident?
incidentsschema
69
A faulty transformation writes negative quantities into 28 million inventory rows across six hours; downstream replenishment starts generating bad orders, so what do you do?
70
An SCD2 customer dimension contains 640,000 overlapping effective-date ranges after a partial rerun, causing orders to join twice; how do you repair it?
joins
71
After adding a new CRM, a Data Vault hub shows 2.3 million duplicate customers because the new source normalizes business keys differently; what is your response?
normalizationbusiness-keyssecrets
72
A Data Vault satellite load generates a new version for 92% of rows every night although source changes average 3%; how do you diagnose it?
secretsdata-vault
73
A Data Vault PIT table is eight hours stale after its incremental job skipped late satellite records, and 70 finance marts show old account status; how do you recover?
secretsdata-vault
74
A dbt partial refresh updates only 4 of 11 upstream partitions, leaving a monthly margin model internally inconsistent by $2.7 million; what do you do?
dbtpartitioningcss
75
A producer changes order_id from string to integer and 23 dbt models fail their contracts two hours before executive reporting; how do you respond?
dbt
76
Airflow has 18,000 queued tasks after the scheduler database slowed down, and the 07:00 reporting SLA is 90 minutes away; how do you prioritize recovery?
airflowdatabaseprioritization
77
A team launches a 180-day Airflow backfill that creates 240,000 tasks and starves daily pipelines across eight domains; what do you do?
airflowci-cdbackfill
78
Finance reports $84.2 million in quarterly revenue while the product semantic layer reports $87.9 million, five hours before earnings review; how do you lead the resolution?
semantic-layer
79
A checkout team removes a nullable promotion field without notice, breaking 31 consumers across four domains during a campaign handling 22,000 orders per minute; what do you do?
80
A regulatory report exposes a wrong total, but column lineage stops at a stored procedure used by 46 downstream models; the filing deadline is tomorrow, so what do you do?
schemastored-proceduresestimation
81
A critical customer table has no catalog owner, 17 stale definitions, and an upstream team plans to delete it in 48 hours; how do you handle the governance failure?
data-governancesoft-skills
82
A marketing export accidentally includes email and birth date for 1.8 million users and remains accessible for 37 minutes; what are your first actions?
83
A Snowflake row access policy regression lets 14 regional managers query records from all 26 regions for 22 minutes; how do you respond?
queriessnowflake
84
A GDPR deletion audit finds that 4,200 users removed from operational databases still exist in Iceberg snapshots and three derived marts after 45 days; what do you do?
databasesnapshotgdpr
85
An MDM model falsely merges 62,000 households after a threshold change, causing loyalty points to move between customers; how do you recover?
86
A data mesh domain publishes inventory data six hours late for the fourth week, violating a 30-minute product SLA used by 12 teams; how do you intervene?
data-mesh
87
Cross-region warehouse replication falls 9 hours behind during a network degradation, and the secondary region is the source for regulatory reporting; what do you do?
replicationwarehouse
88
A quarterly DR test fails because the lake catalog restores in 18 minutes but 27% of table paths point to unavailable regional storage; the RTO is 60 minutes, so what do you change?
89
A multi-cloud pipeline begins moving 2.4 PB per month across providers and egress cost rises to $190,000, six times forecast; how do you contain it?
ci-cd
90
Your shared warehouse is at 92% compute capacity during business hours, demand grows 8% monthly, and procurement needs 10 weeks for expansion; what decision do you make this week?
procurementcapacitywarehouse
91
During an on-premises to cloud cutover, cloud row counts are 0.4% lower across 18 billion transactions and the shutdown window ends in 70 minutes; do you proceed?
transactions
92
After a merger, two customer masters assign the same national ID to 11,000 conflicting legal entities, and sales wants a unified view in two weeks; what do you do?
93
Your managed data catalog vendor is unavailable for 11 hours, blocking schema approvals for 140 pipelines while runtime data remains healthy; how do you operate?
data-catalogschemaci-cd
94
A vendor offers a proprietary streaming feature that cuts latency from 8 minutes to 20 seconds but would bind 260 pipelines to its format before a nine-month launch deadline; do you approve it?
procurementlatencystreaming
95
A streaming balance view differs from the nightly ledger by $1.3 million across 0.06% of accounts after a retry-policy change; how do you reconcile it?
resiliencestreaming
96
A source silently stops sending one country code, removing 9% of daily orders, but no data-quality alert fires for 14 hours; what do you change?
alerting
97
A new observability rollout sends 6,800 data-quality alerts per day, 97% of which self-resolve, and on-call misses a real payroll defect; how do you fix the system?
observabilityalertingsystem-design
98
In a review two days before launch, you find a Python ingestion service loading 600 million rows with string-built SQL, no transaction boundary, and no idempotency; the team says tests pass, so what do you decide?
sqltransactionsidempotency
99
A middle data architect approves a partition redesign that makes a 90 TB table unqueryable for 40 minutes; how do you mentor them while restoring service?
mentoringpartitioning
100
The CFO asks you to skip reconciliation and launch a migrated revenue platform by quarter close, accepting up to a 1% mismatch on $420 million; what do you say and do?
react

Questions

1A retailer has 180 source systems, 14 business units, 6,000 reports, and a $12 million three-year modernization budget. How would you define the enterprise data target state and sequence capabilities?

2A manufacturer stores 4 PB of sensor files, runs 20,000 SQL dashboards, and trains 80 models, with a 15-minute analytics SLA. Would you choose a warehouse, a lakehouse, or both?

3A financial company has 900 analysts, 12,000 governed SQL queries per hour, limited Spark skills, and a requirement to launch in six months. Would you select Snowflake or Databricks?

4A media company ingests 25 TB daily, runs 300 Spark pipelines, trains recommendation models on GPUs, and needs streaming features within 60 seconds. Would you select Snowflake or Databricks?

5A company must expose 2 PB of lake tables to Spark, Trino, Flink, and two cloud vendors without vendor lock-in. Would you standardize on Delta Lake or Apache Iceberg?

6Thirty teams publish 600 tables through bronze, silver, and gold layers, and consumers require a 99.5% freshness SLO. What contracts would you put between the medallion layers?

7An e-commerce platform processes 120,000 events per second, needs fraud features in 3 seconds and finance totals by 06:00, while retaining 7 years of data. Design the real-time plus batch platform.

8A logistics company has 40 event types, 15-minute late arrivals, a 5-second tracking SLA, and only eight platform engineers. Would you choose Lambda or Kappa architecture?

9You must capture changes from 220 PostgreSQL and Oracle databases at 70,000 rows per second with under 2-minute freshness and no source downtime. What CDC architecture would you use?

10A Kafka platform carries 300,000 events per second for 25 domains, and per-customer ordering must hold while no tenant may consume over 5% of capacity. How would you design topics and partitions?

11A payments stream runs at 45,000 events per second and must produce one ledger entry per transaction with a 4-second p99 SLA. What Kafka and Flink consistency contract would you design?

12Spark compacts 80 TB of Iceberg tables nightly while Flink writes 20,000 updates per second, and readers require snapshot-consistent results. How would you coordinate the topology?

13A bank operates in the EU, US, and Singapore, processes 8 TB daily, and forbids raw customer data from leaving its legal region. Design the cross-region architecture.

14An insurer must run analytics across AWS and Azure, move no more than 50 TB per month between clouds, and recover from one provider loss within 8 hours. What multi-cloud design would you choose?

15A marketplace needs sub-20 ms profile reads, ad hoc analysis over 3 PB, full-text search across 500 million listings, and an immutable audit trail. How would you choose storage technologies?

16A global distributor must integrate 140 volatile sources, preserve history for 10 years, and onboard acquisitions within 30 days. How would you design a Data Vault 2.0 core?

17The Data Vault core contains 12 years of sales history, but 1,200 analysts need daily revenue queries under 5 seconds. How would you build dimensional marts without creating a second truth?

18A company has 18 product teams spanning commerce, fulfillment, finance, and customer support, with shared customer and order data. How would you draw data mesh domain boundaries?

19Eight domains publish 240 data products to 600 consumers, and breaking changes require 90 days notice. What would a platform-enforced data product contract contain?

20Twelve domains operate independently, but regulators require one PII classification policy and audit evidence within 48 hours. How would you implement federated governance?

Locked questions