Skip to content
Tech Interview Prep
Technical interview guide

Data Catalog & Lineage

Making data discoverable and traceable: what a dataset means, where it came from, and what depends on it.

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

Scope: W3C DCAT 3 and PROV-O, OpenLineage 1.53 object model/facets, Google Cloud Knowledge Catalog lineage, dbt manifest artifacts, Apache Atlas lineage API, and UK metadata guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-3A data platform team's catalog dashboard reports 98% completeness: nearly every registered table has an owner, a description and a freshness field filled in. A reviewer asks what that number actually establishes. What does it prove?
MCQ-1A platform team runs a catalog over 4,000 Snowflake tables. A nightly crawl indexes every table's schema and stats, finishing all 4,000 by 02:00; a classifier tags PII columns at 94% precision; 40% of tables have no written description. Search returns a plausible match for 68% of queries, yet 61% of data scientists still ask in Slack which of three similarly named tables feeds the finance dashboard, since the catalog gives no way to tell. Which change is most likely to cut those Slack questions?
MCQ-2A platform team's nightly catalog job harvests table and column names, data types, nullability, row counts and query-log lineage from the warehouse. Analysts still file about 40 tickets a month asking which of two similarly named revenue columns is the certified figure for the board pack and who must approve its use in external reporting. Which catalog change addresses those tickets?
MCQ-4A catalog team loads four resources: ridership-2023.csv, ridership-2023.parquet, ridership-2023.jsonl and station-locations.geojson, and enters each one as its own dataset. Search now shows four entries for what users think of as two things, and the format facet, which is built from distributions, is empty. Station locations are owned by a different team and refreshed weekly; ridership is refreshed monthly. Which modeling change fixes the catalog?
MCQ-5A catalog keys every asset by its fully-qualified display name. Last quarter 1,200 tables were renamed or moved between warehouses; the lineage graph lost 340 edges and 90 column tags were re-attached to the wrong table after a new table reused a retired name. The team replaces the key with an immutable asset ID assigned at creation, keeping display names and aliases as mutable attributes. A table is then renamed from raw.events to staging.events_v2. What does the lineage query for that table return?
MCQ-24A catalog harvests 40 source systems nightly. Each run asks every source's metadata API for assets modified since the previous run and writes whatever comes back. After a month, 2,160 of 12,000 tables (18%) have no observation newer than 30 days, and all of them sit in 6 of the 40 sources. Within those six sources, every stale table has had an unchanged schema all month, and every table that does have a fresh observation had at least one schema change in that window. All 280 runs in the last week logged HTTP 200 with no errors, and the six source teams confirm the tables still exist and are queried daily. What is the most likely cause?
MCQ-6A data catalog's nightly index job fills a "sample rows" column by running `SELECT * FROM <table> LIMIT 1` against each of 12,000 warehouse tables under one catalog service account. It takes 40 minutes, fails on the 300 tables where that account lacks SELECT, and those 300 tables never appear in search, so analysts cannot find tables they are allowed to query. Security policy says the catalog account must not hold read access to table contents. What should the team change?
MCQ-17A bank is deciding whether to give a third-party analytics vendor read access to its lineage catalog. The catalog holds dataset names, column names, owner emails, parameterized job SQL with no literal values, and Git paths; it holds no row data, and the vendor gets no warehouse or repository credentials. The vendor's contract lets it build a competing product. Which property of the catalog is the reason the security team blocks the access?
MCQ-19Your warehouse's lineage graph is: `raw_events → stg_events → fct_sessions`; `raw_events → stg_devices`; `stg_devices → fct_sessions`; `fct_sessions → dash_sessions_daily`; `fct_sessions → ml_features → model_churn_scores`. At 02:10 the `stg_events` job fails and writes no rows for the 02:00 partition. The on-call engineer runs a forward (downstream) lineage query rooted at `stg_events` to scope the incident. Which set of assets does that query return?
MCQ-20A nightly job builds `rev` for the finance dashboard. The lineage graph records these edges, parent → child: - `orders` → `stg_o` - `returns` → `stg_r` - `stg_o` → `fct` - `stg_r` → `fct` - `fct` → `rev` - `fct` → `cost` - `rev` → `dash` - `rev` → `report` - `rev` → `alert` - `rev` → `email` - `rev` → `slack` - `promos` → `dim_p` - `dim_p` → `dash` This week's revenue is 12% below last week's and the engineer wants to bound the search for the cause. Which set of tables does backward lineage from `rev` return?
MCQ-7Your team is dropping the `email` column from `dim_customer` in the analytics warehouse. The catalog holds dataset-level lineage only. `dim_customer` has 7 direct downstream tables: 2 select `email`, and the other 5 select only `customer_id` and `name`. Those 5 fan out to 23 further tables, and the 2 that read `email` have no downstream consumers. How many tables does the lineage graph flag as affected, and how many actually break?
MCQ-11A team plans to drop the column `billing_zip` from `stg_orders`. Dataset-level lineage lists 30 downstream models that read `stg_orders`. Column-level lineage shows that 2 of those models select `billing_zip` and the other 28 select only `order_id` and `order_date`. What does the column-level view change about the migration scope?
MCQ-18You are renaming `orders.customer_id` in the warehouse. The SQL-parser lineage graph lists 12 downstream tables. A 90-day query-log scan finds 31 distinct consumer principals reading the column; 9 of them never appear in the graph because they build SQL at runtime. Three of the 12 lineage tables have no registered owner. The change window is Friday. What should the impact analysis conclude?
MCQ-12A team runs a static SQL parser over a dbt project to build column-level lineage; the parser reads the compiled SQL and the warehouse catalog. Which lineage will it still miss?
MCQ-21A catalog builds column-level lineage by parsing the warehouse query log. For `analytics.orders`, the parser resolves 12 downstream column references across 4 tables. One upstream job, `load_orders`, assembles its query from a Jinja template whose source table name is read at runtime from a config table, so the parser emits no edge for it. An engineer runs impact analysis before dropping `orders.legacy_id`. What should the catalog return?
MCQ-13A nightly Spark job is instrumented with an OpenLineage emitter that sends START, COMPLETE and FAIL events to a catalog. A column-level dependency on a table read only when a feature flag is enabled never appears in the graph, and one job's outputs are missing for a full week even though the job ran every night. What does runtime capture actually miss here?
MCQ-25A payments team is about to drop a column from the `orders` table. The catalog's lineage view for `orders` lists three downstream consumers, so they drop it. Two days later a settlement job fails: it reads the column through a view the catalog never listed. The catalog builds lineage nightly by matching column names across the warehouse, and the settlement view aliases the column (`txn_amt` for `amount`). The table page renders every edge in one Lineage list, with no indication of where each edge came from. Which change would most reduce the chance of the same miss?
MCQ-22A lineage service derives column-level edges two ways: static SQL parsing for jobs it can parse, and runtime query logs for the rest. The logs record table reads and writes, so edges from them are table-level only. Last quarter a new orchestrator was onboarded; its jobs are dynamically generated SQL that the parser cannot parse. Total edges in the graph rose from 410k to 533k. Dropping a column on `orders` then flagged 3 downstream tables, but a manual audit found 11, and none of the 8 missed tables had any column-level edge from `orders` in the graph. A sample of 200 edges checked against manually verified ground truth found 174 correct. The critical-data list held the same 40 tables before and after; 37 had at least one column-level upstream edge before the onboarding, 24 after. Which reading of these numbers is right?
MCQ-9A nightly Spark job emits OpenLineage events to a lineage backend. Its final task emits a COMPLETE event for run R at 02:14, naming `warehouse.orders` as an output dataset. At 02:20 the downstream job that reads `warehouse.orders` fails: the table exists but still holds yesterday's partition. The lineage UI shows run R as COMPLETE. What does the COMPLETE event establish about run R?
MCQ-10A lineage service consumes OpenLineage events from the `orders_etl` job and publishes a dataset edge `orders_etl -> analytics.daily_orders` for every START event it sees. Last night the job ran 200 times: 185 runs emitted COMPLETE, 9 emitted FAIL after the write step raised, 4 emitted ABORT while the write was in flight, and 2 emitted COMPLETE after writing zero rows because the source partition was empty. The service published 200 edges. How many of those edges are wrong?
MCQ-8A lineage service ingests OpenLineage RunEvents from a nightly Airflow DAG. The task `etl.orders` runs once per night and emits a START and a COMPLETE event each night. After three nights the UI shows a single node for `etl.orders` with status COMPLETE, and its run history lists one entry. The ingestion code stores each event under the key `namespace + jobName`, overwriting any existing record. Which change makes the run history show three runs?
MCQ-14A lineage collector consumes OpenLineage RunEvents from a Kafka topic with at-least-once delivery. In one day it stores 240,000 rows for 120,000 distinct runIds; the producer re-sends an event when its 30 s publish timeout fires. The graph for `etl_orders` lists its input dataset twice, and a run that emitted FAIL at 10:02 then COMPLETE at 10:05 still shows FAIL. A retry of the 10:02 FAIL arrived at 10:07. The table has an auto-increment primary key and one row per received event. Which change removes both symptoms?
MCQ-16A Spark job writes one output dataset. Your compliance team's extractor and the platform team's extractor both attach a custom dataset facet to that dataset, and both named theirs `piiColumns`: `_producer` `https://acme.example/compliance` for one, `https://acme.example/platform` for the other. Both are added to the same output dataset's `facets` before the RunEvent is sent, and the downstream consumer reads `facets` into a map keyed by facet name. What does that consumer end up with for `piiColumns`?
MCQ-15A lineage collector writes one row per run into a metadata store. A row holds run_id, run_state, the run's newest event_time, and the read and write edges of the last event applied. Run R7 (job etl_orders) is stored as SUCCESS with event_time 12:04:10, write analytics.orders_daily v12, and no read of raw.clicks. A consumer-group rebalance re-delivers records the collector had already consumed, including R7's STARTED event at 12:06:00: event_time 12:03:02, read raw.clicks v3, no write. What should the collector do with that re-delivered event?
MCQ-23A catalog asset `analytics.orders` has a 30-day deprecation countdown set by its owner, ending 14 March. The catalog's deprecation workflow runs nightly; under the catalog's policy it may read lineage and send notifications, but only the asset owner can change the countdown or the asset's status. Lineage shows `settlement_daily` and `finance_export` each read `analytics.orders` within the last 24 hours. What should the workflow do before the countdown ends?