Menu

Data Analyst / Engineer · path test

Data Analyst / Engineer: path test

40 questions from every module of the path, set in new situations. Pass it to complete the path's tests.

40 questions · 2.5 marks each · 100 marks in all · pass mark 70% · your answers are saved as you go ·

Path projects: the test checks your understanding. The capstone projects show you can build, debug and explain them, so do them alongside the test. See the modules →

Question 1 of 40 · DT1A coaching institute exports fees with an amount column holding values such as "₹1,250", "1250.00" and "—". In pandas,Change the column's number format in the spreadsheet program and export the whole file againRemove the symbol and commas, convert with pd.to_numeric, and count the failuresUse astype(float), which skips the rupee signDrop each row that contains a comma so that only the clean numbers are left

Why each option is right or wrong

“Change the column's number format in the spreadsheet program and export the whole file again” (Incorrect): The export still writes text, and the next file arrives the same way; the cleaning belongs in code.
“Use astype(float), which skips the rupee sign” (Incorrect): astype(float) raises an error on "₹1,250"; it doesn't skip symbols.
“Drop each row that contains a comma so that only the clean numbers are left” (Incorrect): That throws away real fees, so the total is wrong in a different way.
“Remove the symbol and commas, convert with pd.to_numeric, and count the failures” is correct.
The column is text, so "sum" joins the strings. Cleaning the text and converting it with errors="coerce" turns "—" into a missing value you can count and report.
- Store: money as whole paise, not floats.
- Report: how many values didn't convert.

📖 Revise: Lesson 3: Cleaning text, numbers and money

Question ID path-test-data-q1 · Report a mistake in this question

Question 2 of 40 · DT1A city bus operator records ticket times in UTC. The daily report groups tickets by date, and tickets sold between midnight andThe bus machines' clocks run slow, so the operator should replace themDates can't be taken from times that have minutes in themThe date is taken in UTC; convert times to Asia/Kolkata firstThe report should round each time to the nearest hour before grouping

Why each option is right or wrong

“The bus machines' clocks run slow, so the operator should replace them” (Incorrect): The times are correct UTC times; the report reads them in the wrong zone.
“Dates can't be taken from times that have minutes in them” (Incorrect): pandas takes a date from any timestamp; the zone is the issue.
“The report should round each time to the nearest hour before grouping” (Incorrect): Rounding moves some tickets but doesn't fix the five-and-a-half-hour offset.
“The date is taken in UTC; convert times to Asia/Kolkata first” is correct.
Midnight to 05:30 in India is still the previous day in UTC. Converting with tz_convert("Asia/Kolkata") before .dt.date puts each ticket on its local day.
- Keep: stored times in UTC, with the zone marked.
- Check: count tickets on the first hour of a day before and after.

📖 Revise: Lesson 4: Dates, times and time zones

Question ID path-test-data-q2 · Report a mistake in this question

Question 3 of 40 · DT1A library joins 5,000 loans to its members table on member_id and gets 5,600 rows. Total fines in the report go up by about 12%.pandas adds rows when a join column has mixed upper and lower caseThe loans table must be sorted by member_id before any joinA left join adds an empty row for each loan that has no memberSome member IDs appear twice in the members table, doubling loans

Why each option is right or wrong

“pandas adds rows when a join column has mixed upper and lower case” (Incorrect): Case differences make rows fail to match; they don't add rows.
“The loans table must be sorted by member_id before any join” (Incorrect): Order doesn't change how many rows a join produces.
“A left join adds an empty row for each loan that has no member” (Incorrect): A left join keeps unmatched loans as one row each; it doesn't add extras.
“Some member IDs appear twice in the members table, doubling loans” is correct.
Each loan matches every copy of its member, so those loans are counted more than once.
- Check: members["member_id"].is_unique.
- Guard: merge(..., validate="many_to_one") raises an error if the right side has repeats.

📖 Revise: Lesson 5: Joining and reshaping without surprises

Question ID path-test-data-q3 · Report a mistake in this question

Question 4 of 40 · DT1A gym chain sends a sign-up export and asks for "a quick analysis". Before writing any cleaning code, what's the most usefulAgree the question and what one row of the export stands forConvert the file to Parquet so that later steps run fasterRemove all columns with any missing values to keep the data tidyBuild the dashboard first, then work back to the data it needs

Why each option is right or wrong

“Convert the file to Parquet so that later steps run faster” (Incorrect): Speed isn't the problem yet; you don't know what the data is for.
“Remove all columns with any missing values to keep the data tidy” (Incorrect): Missing values may matter to the question; dropping them first hides information.
“Build the dashboard first, then work back to the data it needs” (Incorrect): Without the question and the grain, the dashboard may show the wrong measure.
“Agree the question and what one row of the export stands for” is correct.
Knowing the question ("which branches lose members after a month?") and the grain (one row per sign-up, or per member?) decides which columns matter and what "clean" means.
- Write down: the question, the grain, the expected row count.
- Then: look at the raw file before changing it.

📖 Revise: Lesson 1: Raw data and the question behind it

Question ID path-test-data-q4 · Report a mistake in this question

Question 5 of 40 · DT2A courier company's nightly load reports: 1,000 rows read, 960 loaded, 25 quarantined. The dashboard updated as usual. WhatNothing; 96% loaded is within normal quality for real dataLoad the 25 quarantined rows as they are, so the totals are completeFind the 15 rows that were neither loaded nor quarantined, and whyRe-run the load until the loaded count reaches 1,000 rows

Why each option is right or wrong

“Nothing; 96% loaded is within normal quality for real data” (Incorrect): The share isn't the issue; some rows are unaccounted for.
“Load the 25 quarantined rows as they are, so the totals are complete” (Incorrect): Those rows failed the rules for a reason; loading them adds bad data.
“Re-run the load until the loaded count reaches 1,000 rows” (Incorrect): Re-running doesn't explain the gap, and may add repeats.
“Find the 15 rows that were neither loaded nor quarantined, and why” is correct.
Rows read should equal rows loaded plus quarantined plus any rows removed on purpose. 15 are missing from that sum, so something dropped them without saying why.
- Rule: stop or warn when the counts don't add up.
- Log: each count in a load log.

📖 Revise: Lesson 4: Counts that add up

Question ID path-test-data-q5 · Report a mistake in this question

Question 6 of 40 · DT2A pharmacy chain's daily stock file sometimes contains negative quantities. Today the whole load stopped on the first one, andQuarantine the failing rows with a reason and load the restTurn negative quantities into zero, so the load can carry onSkip validation on busy days so that the file loads on timeAsk each shop to check its own numbers by hand before uploading them

Why each option is right or wrong

“Turn negative quantities into zero, so the load can carry on” (Incorrect): That hides a real problem, such as returns or a data-entry error.
“Skip validation on busy days so that the file loads on time” (Incorrect): Bad data would then reach the managers on exactly the days it matters most.
“Ask each shop to check its own numbers by hand before uploading them” (Incorrect): Manual checks are slow and inconsistent; rules in the pipeline catch the same problems each day.
“Quarantine the failing rows with a reason and load the rest” is correct.
The good rows reach managers on time, and the bad ones are kept, labelled and counted, so someone can fix them at the source.
- Keep: the reason, the source file and the row.
- Report: how many rows each rule caught.

📖 Revise: Lesson 3: Validation rules and quarantine

Question ID path-test-data-q6 · Report a mistake in this question

Question 7 of 40 · DT2A hotel group hands you a bookings dataset you haven't seen. Which first look tells you most about its quality?Draw the monthly revenue chart, since a strange chart shows the problemsRead the first 20 rows carefully to judge the whole datasetCheck that the file opens without errors in a spreadsheet programCount rows, missing values, distinct values and repeated keys

Why each option is right or wrong

“Draw the monthly revenue chart, since a strange chart shows the problems” (Incorrect): A chart hides missing values, repeats and odd categories.
“Read the first 20 rows carefully to judge the whole dataset” (Incorrect): The first rows may be the cleanest; problems are often further down.
“Check that the file opens without errors in a spreadsheet program” (Incorrect): Opening without errors says little about what's inside.
“Count rows, missing values, distinct values and repeated keys” is correct.
A profile shows the size, gaps, odd categories, ranges and duplicates across all rows, not just a sample.
- Look at: min and max of dates and amounts.
- Ask: is the expected key unique?

📖 Revise: Lesson 2: Profiling a new dataset

Question ID path-test-data-q7 · Report a mistake in this question

Question 8 of 40 · DT2You've written a function that cleans phone numbers for a clinic's patient list. How do you test it well?Run it on the full patient list and check that the row count is unchangedSmall made-up inputs with known answers, including odd casesAsk a colleague to read the code and confirm it looks rightCopy a few real patient records into the test file for realism

Why each option is right or wrong

“Run it on the full patient list and check that the row count is unchanged” (Incorrect): The count can be right while the values are wrong.
“Ask a colleague to read the code and confirm it looks right” (Incorrect): A review helps, but it doesn't run the code on the tricky cases.
“Copy a few real patient records into the test file for realism” (Incorrect): Real patient data doesn't belong in test files; synthetic cases test the same rules.
“Small made-up inputs with known answers, including odd cases” is correct.
Hand-made inputs such as an empty value, a number with spaces, a "+91" prefix and a too-short number show exactly what the function does.
- Tool: pytest, with synthetic data only.
- Keep: a test for each bug you fix.

📖 Revise: Lesson 5: Testing data code

Question ID path-test-data-q8 · Report a mistake in this question

Question 9 of 40 · DT3An electricity board's pipeline appends each day's meter file to a table. After a crash, someone re-ran yesterday, andAdd a warning in the README that says not to run the job twiceLoad each file into a new table with the date in its nameRemove duplicate rows from the whole table once a monthReplace that file's rows in one transaction when it's loaded

Why each option is right or wrong

“Add a warning in the README that says not to run the job twice” (Incorrect): People and schedulers re-run jobs; the design has to handle it.
“Load each file into a new table with the date in its name” (Incorrect): That avoids doubling but scatters the data and makes queries harder.
“Remove duplicate rows from the whole table once a month” (Incorrect): Reports are wrong for up to a month, and true repeats in the source get removed too.
“Replace that file's rows in one transaction when it's loaded” is correct.
Deleting the rows from that source file and inserting the new ones in one transaction makes a re-run give the same result as one run.
- Name: an idempotent load.
- Test: load the same file twice and compare counts.

📖 Revise: Lesson 2: Safe to run again (idempotent loads)

Question ID path-test-data-q9 · Report a mistake in this question

Question 10 of 40 · DT3A food-delivery pipeline loads only "yesterday's orders" each morning. Refunds are often recorded two days after the order, soRun the load once a week instead, after most refunds are inEach run reloads the last three days of orders, replacing themAsk the refund team to record refunds on the same day as the orderMark every order as "pending" until its delivery partner confirms the payment

Why each option is right or wrong

“Run the load once a week instead, after most refunds are in” (Incorrect): Reports then wait a week, and late refunds can still arrive after that.
“Ask the refund team to record refunds on the same day as the order” (Incorrect): That may not be possible, and the pipeline should handle late data anyway.
“Mark every order as "pending" until its delivery partner confirms the payment” (Incorrect): That changes the business meaning and doesn't load the refunds.
“Each run reloads the last three days of orders, replacing them” is correct.
A look-back window picks up late changes to recent orders, and an idempotent load replaces the earlier versions.
- Choose: the window from how late changes really arrive.
- Backfill: run longer windows when needed.

📖 Revise: Lesson 3: Incremental loads, backfills and late data

Question ID path-test-data-q10 · Report a mistake in this question

Question 11 of 40 · DT3A school's attendance pipeline failed at 2 a.m. Nobody knew until teachers saw yesterday's figures at 10 a.m. What wouldA failing exit code, a load log, and an alert to the data ownerRunning the pipeline at 9 a.m., when staff are already at workWriting the pipeline in a faster language so it finishes earlierRunning it three times each night in the hope that one run works

Why each option is right or wrong

“Running the pipeline at 9 a.m., when staff are already at work” (Incorrect): Failures would be noticed sooner but reports would be later; it doesn't detect failure.
“Writing the pipeline in a faster language so it finishes earlier” (Incorrect): Speed doesn't tell anyone that it failed.
“Running it three times each night in the hope that one run works” (Incorrect): Repeats don't report failure and can hide a real problem.
“A failing exit code, a load log, and an alert to the data owner” is correct.
A job that exits with an error, writes what happened to a load log and alerts someone turns a silent failure into a morning fix.
- Also: a freshness check on the latest load time.
- Show: "data as of" on the report.

📖 Revise: Lesson 5: When things go wrong

Question ID path-test-data-q11 · Report a mistake in this question

Question 12 of 40 · DT3A retailer's data work has three steps: extract, clean, build reports. Each is a separate scheduled job at 1:00, 1:30 and 2:00.Move the three times further apart so each step has more timeCombine all three steps into one script with no loggingOne workflow with clear dependencies between the three stepsAsk the report users to wait until 3:00 before opening reports

Why each option is right or wrong

“Move the three times further apart so each step has more time” (Incorrect): That helps until a night is slow enough again, and it delays every report.
“Combine all three steps into one script with no logging” (Incorrect): One script keeps the order but loses visibility into which step failed.
“Ask the report users to wait until 3:00 before opening reports” (Incorrect): The data can still be half-cleaned at 3:00 on a slow enough night.
“One workflow with clear dependencies between the three steps” is correct.
An orchestrator starts each step when the one before it succeeds, and stops if it fails, so reports aren't built from half-finished data.
- Examples: Airflow, Dagster, Prefect, or a CI workflow for small jobs.
- Plus: retries and alerts per step.

📖 Revise: Lesson 4: Scheduling and orchestration

Question ID path-test-data-q12 · Report a mistake in this question

Question 13 of 40 · DT4A bank needs to flag suspicious card payments within seconds, and also sends a monthly spending summary to each customer. WhichStreaming for the payment checks, batch for the summaryBatch processing for both, run every hour to keep the systems simpleStreaming for both, since streaming is the more modern approachA dashboard that staff refresh by hand when they suspect fraud

Why each option is right or wrong

“Batch processing for both, run every hour to keep the systems simple” (Incorrect): Hourly is far too slow for stopping a fraudulent payment.
“Streaming for both, since streaming is the more modern approach” (Incorrect): The summary gains nothing from seconds-old data and costs more to run.
“A dashboard that staff refresh by hand when they suspect fraud” (Incorrect): People can't watch every payment within seconds.
“Streaming for the payment checks, batch for the summary” is correct.
The fraud check needs answers in seconds, so events are processed as they arrive. The monthly summary can wait, so a cheaper scheduled batch job fits.
- Rule of thumb: choose by how soon the answer is needed.
- Cost: streams run all the time.

📖 Revise: Lesson 1: Batch or stream?

Question ID path-test-data-q13 · Report a mistake in this question

Question 14 of 40 · DT4A ride-hailing app counts rides per five-minute window. Phones in poor network areas send events up to two minutes late, and thoseCount rides by the time the server receives them, so lateness doesn't matterThrow away events that arrive late, since they're a small shareUse one-hour windows so that late events still fit inside themWindow by event time, with a watermark that allows late events

Why each option is right or wrong

“Count rides by the time the server receives them, so lateness doesn't matter” (Incorrect): Late rides then land in the wrong window, so the counts are still wrong.
“Throw away events that arrive late, since they're a small share” (Incorrect): That undercounts areas with poor networks, which may be the areas that matter.
“Use one-hour windows so that late events still fit inside them” (Incorrect): Larger windows lose the detail the counts were meant to give.
“Window by event time, with a watermark that allows late events” is correct.
Grouping by when the ride happened, and keeping windows open a little after they end, lets late events join the right window.
- Set: the allowed lateness from measured delays.
- Count: events that arrive later than allowed.

📖 Revise: Lesson 3: Processing streams: windows and late events

Question ID path-test-data-q14 · Report a mistake in this question

Question 15 of 40 · DT4A weather research group has 40 GB of sensor CSV files. Most queries read 3 of the 60 columns for one month. Queries are slow.Zip the CSV files to make them smaller on diskColumnar Parquet files, partitioned by month of readingSplit each CSV file into smaller files of 1,000 rows eachLoad all 40 GB into a spreadsheet so researchers can filter it

Why each option is right or wrong

“Zip the CSV files to make them smaller on disk” (Incorrect): Zipped CSV still has to be read and unzipped in full.
“Split each CSV file into smaller files of 1,000 rows each” (Incorrect): Many small files add overhead and still read every column.
“Load all 40 GB into a spreadsheet so researchers can filter it” (Incorrect): Spreadsheets have row limits and are slow at this size.
“Columnar Parquet files, partitioned by month of reading” is correct.
Parquet stores each column separately and compresses well, so queries read only the columns they use; partitions let them skip other months.
- Check: compare file size and the data scanned.
- Tools: DuckDB, Spark and cloud warehouses read it.

📖 Revise: Lesson 5: File formats and data lakes

Question ID path-test-data-q15 · Report a mistake in this question

Question 16 of 40 · DT4A start-up's daily data is about 2 GB and its reports run in five minutes on one laptop. Someone proposes a Spark cluster. What'sSpark is needed for any dataset bigger than 1 GBSpark costs less than one machine, so the team should move straight awayOne machine with DuckDB or pandas is likely enough for nowSpark makes the existing reports more accurate than pandas does

Why each option is right or wrong

“Spark is needed for any dataset bigger than 1 GB” (Incorrect): There's no such cut-off; single machines handle several gigabytes well.
“Spark costs less than one machine, so the team should move straight away” (Incorrect): A cluster costs more to run and look after.
“Spark makes the existing reports more accurate than pandas does” (Incorrect): Both compute the same answers; Spark spreads the work.
“One machine with DuckDB or pandas is likely enough for now” is correct.
A cluster adds cost and moving parts. It pays off when data or processing outgrows one machine, which this doesn't yet.
- Watch: run time and memory as data grows.
- Plan: keep code in SQL or dataframes so it can move later.

📖 Revise: Lesson 4: Spark and when you need a cluster

Question ID path-test-data-q16 · Report a mistake in this question

Question 17 of 40 · DT5Analysts at an online bookshop run big reports straight on the shop's live database, and checkout slows down at month end. What'sAsk analysts to run their reports only at night, when sales are lowerCopy the data to a separate warehouse built for analysisGive the live database a bigger server and keep the reports as they areTurn off the reports in the last week of each month

Why each option is right or wrong

“Ask analysts to run their reports only at night, when sales are lower” (Incorrect): It may help, but reports still compete with the shop's database.
“Give the live database a bigger server and keep the reports as they are” (Incorrect): It buys time but keeps the two kinds of work fighting.
“Turn off the reports in the last week of each month” (Incorrect): That removes the reports when they're needed most.
“Copy the data to a separate warehouse built for analysis” is correct.
The live database is built for many small, fast changes; a warehouse is built for large reads across history. Separating them protects checkout.
- Load: on a schedule, with tests.
- Bonus: history and data from other systems in one place.

📖 Revise: Lesson 1: Why analysis needs a warehouse

Question ID path-test-data-q17 · Report a mistake in this question

Question 18 of 40 · DT5A team pays a cloud warehouse by the amount of data each query reads. One analyst runs SELECT * on a 2 TiB events table, filteringAdd LIMIT 100 to the query, which reduces the data readRun the same query at night, when the prices are lowerSelect the needed columns and filter on the partition columnExport the full table to CSV so that it can be filtered locally

Why each option is right or wrong

“Add LIMIT 100 to the query, which reduces the data read” (Incorrect): On many warehouses LIMIT doesn't reduce the data scanned.
“Run the same query at night, when the prices are lower” (Incorrect): On-demand pricing is per amount read, not by time of day.
“Export the full table to CSV so that it can be filtered locally” (Incorrect): Exporting still reads the whole table.
“Select the needed columns and filter on the partition column” is correct.
Columnar warehouses read only the columns named, and a filter on the partition column skips whole partitions, so the query reads much less.
- Illustrative: at $6.25 per TiB (checked 1 Nov 2026), 2 TiB costs about $12.50 a run.
- Check: the dry-run estimate before running.

📖 Revise: Lesson 4: Queries that read less

Question ID path-test-data-q18 · Report a mistake in this question

Question 19 of 40 · DT5An analyst spots a misspelled city name and fixes it by editing the raw orders table directly. What's the problem?Raw data should stay as received; make the fix in stagingThe edit is fine, but it should be done during the night to avoid locksCity names can't be changed once a table has been createdThe correct fix is to delete the misspelled rows from every layer

Why each option is right or wrong

“The edit is fine, but it should be done during the night to avoid locks” (Incorrect): The timing isn't the problem; changing the source copy is.
“City names can't be changed once a table has been created” (Incorrect): Values can be changed; the question is where to change them.
“The correct fix is to delete the misspelled rows from every layer” (Incorrect): Those are real orders; deleting them loses data.
“Raw data should stay as received; make the fix in staging” is correct.
Keeping raw unchanged lets you re-run everything and prove what the source sent. A fix in staging code is reviewed, repeatable and applies to new files too.
- Layers: raw, then staging, then marts.
- Log: the fix in code, not by hand.

📖 Revise: Lesson 2: Layers and schemas

Question ID path-test-data-q19 · Report a mistake in this question

Question 20 of 40 · DT5A marketing team wants weekly order trends by area. The orders data also holds customers' names and phone numbers. How shouldGive the team full access, since they've signed a confidentiality formEmail them a weekly spreadsheet exported from the raw tablesRemove names from the raw table so everyone can use itA role that can read an area mart without personal columns

Why each option is right or wrong

“Give the team full access, since they've signed a confidentiality form” (Incorrect): A form doesn't reduce what they can see, or what leaks if an account is misused.
“Email them a weekly spreadsheet exported from the raw tables” (Incorrect): Spreadsheets spread personal data and go out of date.
“Remove names from the raw table so everyone can use it” (Incorrect): Other work may need the raw data; access should be controlled, not the source changed.
“A role that can read an area mart without personal columns” is correct.
Least privilege gives the team what the job needs and keeps phone numbers and names out of reach.
- Law: India's DPDP Act asks for purpose limits and safeguards.
- Check: list what each role can read.

📖 Revise: Lesson 5: Access, governance and personal data

Question ID path-test-data-q20 · Report a mistake in this question

Question 21 of 40 · DT6A cinema chain's sales table mixes rows for single tickets with rows holding each day's totals. Monthly revenue comes out roughlyAdd a column that marks which rows are totals, and ask analysts to filter it outDivide the monthly revenue by two in the reportKeep only the daily total rows and drop the single ticketsOne declared grain per fact table, such as one row per ticket

Why each option is right or wrong

“Add a column that marks which rows are totals, and ask analysts to filter it out” (Incorrect): That works only while each analyst remembers the filter.
“Divide the monthly revenue by two in the report” (Incorrect): The share of total rows varies, so the result is still wrong.
“Keep only the daily total rows and drop the single tickets” (Incorrect): That loses the detail needed for questions about films and seats.
“One declared grain per fact table, such as one row per ticket” is correct.
Mixing grains counts the same sale twice. Daily totals belong in a separate summary built from the ticket fact.
- Write down: the grain in the model's description.
- Test: the key is unique at that grain.

📖 Revise: Lesson 1: Facts, dimensions and grain

Question ID path-test-data-q21 · Report a mistake in this question

Question 22 of 40 · DT6A shoe retailer's mart has one row per store per day, with margin_pct. A manager averages margin_pct across all stores for thePercentages don't add; compute total margin over total salesPercentages should be rounded first and then averaged across storesThe average is fine as long as each store has the same number of daysA median would fix it because it isn't moved by large stores

Why each option is right or wrong

“Percentages should be rounded first and then averaged across stores” (Incorrect): Rounding adds error and doesn't fix the weighting.
“The average is fine as long as each store has the same number of days” (Incorrect): Stores still differ in sales size, so the weighting is still wrong.
“A median would fix it because it isn't moved by large stores” (Incorrect): A median answers a different question and still isn't the chain's margin.
“Percentages don't add; compute total margin over total sales” is correct.
A small store's day counts as much as a big store's. Summing the margin and the sales first, then dividing, weights each sale fairly.
- Store: the parts (margin, sales), not only the ratio.
- Same for: averages and conversion rates.

📖 Revise: Lesson 3: Measures that add up (and those that don't)

Question ID path-test-data-q22 · Report a mistake in this question

Question 23 of 40 · DT6An insurance company reports claims by the customer's city. When customers move, all their past claims jump to the new city. WhatCopy the city into each claim row and stop updating the customer tableRebuild all past reports and store them as PDFs before each moveKeep a row per address period, with valid-from and valid-to datesAdd columns for city_1, city_2 and city_3 to the customer table

Why each option is right or wrong

“Copy the city into each claim row and stop updating the customer table” (Incorrect): It freezes the city per claim but leaves the customer table wrong for current use.
“Rebuild all past reports and store them as PDFs before each move” (Incorrect): That keeps old reports but not the data needed for new questions.
“Add columns for city_1, city_2 and city_3 to the customer table” (Incorrect): People can move more than three times, and claims can't tell which column to use.
“Keep a row per address period, with valid-from and valid-to dates” is correct.
A type 2 slowly changing dimension keeps each version, so each claim joins to the city that was true at the time.
- Join: on the customer and the claim date between the dates.
- Current: the row with an open end date.

📖 Revise: Lesson 4: History that remembers (slowly changing dimensions)

Question ID path-test-data-q23 · Report a mistake in this question

Question 24 of 40 · DT6A telecom mart uses customers' mobile numbers as the key, and the mart is shared with an outside research partner. What'sThe mobile number with its last four digits replaced by starsA salted hash of the number, with the salt kept outside the codeThe mobile number written backwards, so it doesn't look like a numberA plain SHA-256 of the mobile number with nothing added to it

Why each option is right or wrong

“The mobile number with its last four digits replaced by stars” (Incorrect): Six visible digits plus area data can still identify people, and stars can collide.
“The mobile number written backwards, so it doesn't look like a number” (Incorrect): Anyone can reverse it.
“A plain SHA-256 of the mobile number with nothing added to it” (Incorrect): Mobile numbers are few enough to hash them all and match them back.
“A salted hash of the number, with the salt kept outside the code” is correct.
A keyed hash gives a stable key for joins without exposing the number, and without the secret salt nobody can rebuild it from a list of numbers.
- Keep: the salt in an environment variable or secret store.
- Note: this is pseudonymisation, so the data is still personal.

📖 Revise: Lesson 5: Keys, and keys that protect people

Question ID path-test-data-q24 · Report a mistake in this question

Question 25 of 40 · DT7A dbt model selects from analytics_dev.stg_payments. It works on the analyst's laptop but reads stale development data inChange the name to analytics_prod.stg_payments before each production runCreate a view called analytics_dev.stg_payments in production tooUse {{ ref('stg_payments') }}, so dbt picks the schema for each targetCopy the staging SQL into the model so that it doesn't depend on another one

Why each option is right or wrong

“Change the name to analytics_prod.stg_payments before each production run” (Incorrect): Editing names by hand is easy to forget and breaks development.
“Create a view called analytics_dev.stg_payments in production too” (Incorrect): Production then depends on a development name, which is confusing and fragile.
“Copy the staging SQL into the model so that it doesn't depend on another one” (Incorrect): Copies drift apart, and fixes have to be made twice.
“Use {{ ref('stg_payments') }}, so dbt picks the schema for each target” is correct.
ref lets dbt pick the right schema for each target and know the build order.
- Raw tables: use {{ source(...) }}.
- Bonus: lineage in the docs.

📖 Revise: Lesson 1: Models, `ref` and `source`

Question ID path-test-data-q25 · Report a mistake in this question

Question 26 of 40 · DT7A clinic's appointment mart once had duplicate appointment IDs, which doubled the "no-show" rate for a week before anyone noticed.Run a query by hand each Monday that counts duplicate IDs in the martunique and not_null tests on the key in schema.ymlAdd DISTINCT to the final query so duplicates disappear from itDocument the key in the README so analysts know it should be unique

Why each option is right or wrong

“Run a query by hand each Monday that counts duplicate IDs in the mart” (Incorrect): Weekly checks still let a week of wrong numbers through.
“Add DISTINCT to the final query so duplicates disappear from it” (Incorrect): That hides the cause and can merge rows that are really different.
“Document the key in the README so analysts know it should be unique” (Incorrect): Documentation doesn't check anything.
“unique and not_null tests on the key in schema.yml” is correct.
dbt build runs the tests with the models, and a failing test stops the models that depend on it.
- Add: a singular test for business rules.
- Run: in CI and on each scheduled build.

📖 Revise: Lesson 2: Tests that run on every build

Question ID path-test-data-q26 · Report a mistake in this question

Question 27 of 40 · DT7An app-events model is rebuilt from scratch every night and now takes two hours. Old events don't change, but some arrive up to a dayA view, so that nothing needs to be built at nightA table, rebuilt only on Sundays to save time on weekdaysAn ephemeral model, so dbt doesn't store anything at allIncremental with a unique key and a short look-back

Why each option is right or wrong

“A view, so that nothing needs to be built at night” (Incorrect): Each query then reads all events, which is slow and costly.
“A table, rebuilt only on Sundays to save time on weekdays” (Incorrect): Reports are out of date for most of the week.
“An ephemeral model, so dbt doesn't store anything at all” (Incorrect): Ephemeral models are inlined into others and stored nowhere, so the cost moves elsewhere.
“Incremental with a unique key and a short look-back” is correct.
Each run processes recent events only, and the unique key with delete+insert or merge replaces late or changed rows.
- Look-back: at least as long as the lateness.
- Check: --full-refresh now and then matches the incremental totals.

📖 Revise: Lesson 4: Materialisations: view, table, incremental

Question ID path-test-data-q27 · Report a mistake in this question

Question 28 of 40 · DT7Two analysts edit the same revenue model. One change was pushed straight to production and broke three dashboards. What teamPull requests where CI runs dbt build in its own schemaOnly one analyst may change dbt models, to avoid conflictsEach analyst keeps their own copy of the model with their name in itChanges are made in production after 6 p.m., when fewer people look

Why each option is right or wrong

“Only one analyst may change dbt models, to avoid conflicts” (Incorrect): That creates a bottleneck and a single point of failure.
“Each analyst keeps their own copy of the model with their name in it” (Incorrect): Copies drift, and dashboards can't tell which one is right.
“Changes are made in production after 6 p.m., when fewer people look” (Incorrect): Broken dashboards are still broken the next morning.
“Pull requests where CI runs dbt build in its own schema” is correct.
Each change is built and tested away from production, and reviewed, before it's merged.
- Plus: state:modified+ runs only what changed and its children.
- Protect: the main branch.

📖 Revise: Lesson 5: dbt in a team

Question ID path-test-data-q28 · Report a mistake in this question

Question 29 of 40 · DT8A grocery chain wants the top three products by revenue in each store. Which SQL approach fits?ROW_NUMBER() partitioned by store and ordered by revenue, keeping ranks 1 to 3ORDER BY revenue DESC LIMIT 3 on the whole tableGROUP BY store with MAX(revenue) repeated three timesA separate query for each store, joined with UNION ALL

Why each option is right or wrong

“ORDER BY revenue DESC LIMIT 3 on the whole table” (Incorrect): That returns three products overall, not three per store.
“GROUP BY store with MAX(revenue) repeated three times” (Incorrect): MAX gives one value per store, not three products.
“A separate query for each store, joined with UNION ALL” (Incorrect): It works but has to change whenever a store opens.
“ROW_NUMBER() partitioned by store and ordered by revenue, keeping ranks 1 to 3” is correct.
A window function ranks rows within each store without collapsing them, and the outer query keeps the top three.
- Ties: RANK() or DENSE_RANK() if equal revenues should share a place.
- Readable: put the ranking in a CTE.

📖 Revise: Lesson 2: Window functions

Question ID path-test-data-q29 · Report a mistake in this question

Question 30 of 40 · DT8An online course site wants to know how many learners who joined in March were still active in April, May and June. What's thisA funnel: count the learners at each lesson they reachedA running total of learners per month, over all monthsAverage daily active learners, compared month by monthA cohort table: group by joining month, count activity later

Why each option is right or wrong

“A funnel: count the learners at each lesson they reached” (Incorrect): Funnels follow steps in a process, not activity over months.
“A running total of learners per month, over all months” (Incorrect): A running total mixes new and old learners.
“Average daily active learners, compared month by month” (Incorrect): That doesn't follow the March group.
“A cohort table: group by joining month, count activity later” is correct.
Each learner is placed in the month they joined, and later months count how many of that group were active.
- Show: as a share of the cohort's size.
- Careful: recent cohorts haven't had time to reach later months.

📖 Revise: Lesson 3: Cohorts and retention

Question ID path-test-data-q30 · Report a mistake in this question

Question 31 of 40 · DT8A travel site's funnel shows 120% conversion from "viewed hotel" to "booked". The query counts rows in each event table. What'sConversion above 100% is normal on busy holiday weekendsIt counts events, not people who did each step in orderThe tables use different time zones, so the counts can't be comparedBookings should be counted from the payments table only

Why each option is right or wrong

“Conversion above 100% is normal on busy holiday weekends” (Incorrect): A funnel step can't have more people than the step before it.
“The tables use different time zones, so the counts can't be compared” (Incorrect): Zones could shift a few events, but not create 20% more bookings than views.
“Bookings should be counted from the payments table only” (Incorrect): The source may matter, but counting events instead of people is the main error.
“It counts events, not people who did each step in order” is correct.
One person can book twice or view once and book later. Counting distinct users who reached each step, after the step before, gives a real conversion rate.
- Set: a time limit between steps.
- Check: each step is at most the one before.

📖 Revise: Lesson 4: Funnels and conversion

Question ID path-test-data-q31 · Report a mistake in this question

Question 32 of 40 · DT8A query for "orders that weren't cancelled" uses WHERE status <> 'cancelled'. Some orders with no status yet are missing from<> only works on numbers, not on text columnsThe database skips rows that were inserted most recentlyA comparison with NULL is unknown, so those rows are droppedText comparisons depend on the alphabet and capital letters are sorted first

Why each option is right or wrong

“<> only works on numbers, not on text columns” (Incorrect): <> works on text.
“The database skips rows that were inserted most recently” (Incorrect): Recent rows aren't skipped; the comparison is the issue.
“Text comparisons depend on the alphabet and capital letters are sorted first” (Incorrect): Case can matter, but the missing rows have no status at all.
“A comparison with NULL is unknown, so those rows are dropped” is correct.
NULL <> 'cancelled' isn't true, so WHERE leaves the row out.
- Fix: status IS DISTINCT FROM 'cancelled', or COALESCE(status, 'none').
- Ask: should an order with no status be counted at all?

📖 Revise: Lesson 5: Traps that make numbers wrong

Question ID path-test-data-q32 · Report a mistake in this question

Question 33 of 40 · DT9A store manager asks you for "a dashboard with all the numbers". What's the most useful first step?Gather every metric the warehouse has and put them on one pageAsk what decision it supports and what they'd changeCopy a dashboard template from a BI tool's galleryAsk which colours the manager likes for the charts

Why each option is right or wrong

“Gather every metric the warehouse has and put them on one page” (Incorrect): A page with everything is hard to read and supports no particular decision.
“Copy a dashboard template from a BI tool's gallery” (Incorrect): A template answers someone else's questions.
“Ask which colours the manager likes for the charts” (Incorrect): Colours come later; the content has to be right first.
“Ask what decision it supports and what they'd change” is correct.
The decision (for example, how much stock to order) tells you which measures, time range and breakdowns matter, and which to leave out.
- Write down: the question, the measure and how often it's needed.
- Then: a sketch before any build.

📖 Revise: Lesson 1: Start from the decision

Question ID path-test-data-q33 · Report a mistake in this question

Question 34 of 40 · DT9A bar chart compares four branches' revenue. The axis starts at ₹9 lakh, so a branch with ₹10 lakh looks five times bigger than oneStart the bar axis at zero so lengths match valuesUse 3D bars so the differences look more balancedChange the colours so the larger branch looks less strikingSwitch to a pie chart so the axis isn't a problem any more

Why each option is right or wrong

“Use 3D bars so the differences look more balanced” (Incorrect): 3D distorts lengths further.
“Change the colours so the larger branch looks less striking” (Incorrect): Colour doesn't fix the misleading lengths.
“Switch to a pie chart so the axis isn't a problem any more” (Incorrect): Pies make small differences harder to see.
“Start the bar axis at zero so lengths match values” is correct.
Bar length is read as size, so a cut-off axis exaggerates differences. Starting at zero shows the real ratio.
- Small differences: show the numbers, or a dot plot.
- Lines: can start elsewhere if labelled.

📖 Revise: Lesson 3: Honest and accessible charts

Question ID path-test-data-q34 · Report a mistake in this question

Question 35 of 40 · DT9A head of operations wants to see how one delivery-time measure changed over the last 18 months. Which chart fits best?A pie chart with one slice for each monthA table of 18 numbers with the largest shown in redA line chart with the months along the bottomA stacked bar chart with the months stacked on top of each other

Why each option is right or wrong

“A pie chart with one slice for each month” (Incorrect): Slices show parts of a whole, not change over time.
“A table of 18 numbers with the largest shown in red” (Incorrect): The numbers are there, but the trend is hard to see.
“A stacked bar chart with the months stacked on top of each other” (Incorrect): Stacking months makes each month hard to compare.
“A line chart with the months along the bottom” is correct.
A line shows the direction and shape of change over time at a glance.
- Add: a target line if there is one.
- Accessible: a data table and a text summary.

📖 Revise: Lesson 2: Choosing the right chart

Question ID path-test-data-q35 · Report a mistake in this question

Question 36 of 40 · DT9A public dashboard shows customers per pin code. Some pin codes have one or two customers, and a journalist could work out who theyShow those groups in a lighter colour so they draw less attentionAdd a note asking viewers not to identify peopleRemove the dashboard and send the numbers by email insteadHide or merge any group below a set minimum count

Why each option is right or wrong

“Show those groups in a lighter colour so they draw less attention” (Incorrect): The numbers are still there.
“Add a note asking viewers not to identify people” (Incorrect): A note doesn't stop anyone.
“Remove the dashboard and send the numbers by email instead” (Incorrect): Email spreads the same detail further.
“Hide or merge any group below a set minimum count” is correct.
Suppressing small counts, or merging them into "other", stops a chart from pointing at a few people.
- Example: the kit's minimum is 5.
- Check: totals can reveal hidden cells by subtraction.

📖 Revise: Lesson 5: Privacy, freshness and trust on the page

Question ID path-test-data-q36 · Report a mistake in this question

Question 37 of 40 · DT10In the capstone, your dashboard shows ₹3,000 more net revenue than the warehouse mart for the same month. What's the right nextRound both numbers to the nearest lakh so that they matchTrust the dashboard, since it's what the users actually seeNote the difference in the case study and move on to the demoFind the cause with a reconciliation test, then fix it

Why each option is right or wrong

“Round both numbers to the nearest lakh so that they match” (Incorrect): Rounding hides a real difference in the logic.
“Trust the dashboard, since it's what the users actually see” (Incorrect): The mart is the agreed definition; the dashboard has to match it.
“Note the difference in the case study and move on to the demo” (Incorrect): A known mismatch undermines trust in every number on the page.
“Find the cause with a reconciliation test, then fix it” is correct.
A test that compares the dashboard's total with the mart's shows where they differ, such as cancelled or refunded orders, so you can fix the logic.
- Keep: the test, so it fails if it happens again.
- Explain: the cause in the design notes.

📖 Revise: Lesson 4: Project 3: a dashboard people can trust

Question ID path-test-data-q37 · Report a mistake in this question

Question 38 of 40 · DT10While working on Project 2, you notice the salt for customer_key is written in dbt_project.yml, which is in the repository.Keep it, because the repository is private for the courseMove it into a comment so that dbt no longer reads itRead it from an environment variable and change the saltEncode the salt in base64 so it can't be read easily

Why each option is right or wrong

“Keep it, because the repository is private for the course” (Incorrect): Private repositories are shared, copied and sometimes made public.
“Move it into a comment so that dbt no longer reads it” (Incorrect): It's still in the file and the history.
“Encode the salt in base64 so it can't be read easily” (Incorrect): Base64 is reversible by anyone.
“Read it from an environment variable and change the salt” is correct.
The code reads env_var('SABZI_KEY_SALT') so the secret stays out of the repository, and a new salt makes the old one useless.
- Then: rebuild the keys.
- Check: search the history for the old value.

📖 Revise: Lesson 3: Project 2: a warehouse with personal data protected

Question ID path-test-data-q38 · Report a mistake in this question

Question 39 of 40 · DT10Your Project 1 pipeline gives the right totals on a clean month of data. Which extra evidence best shows it can be trusted?Run it twice on a file with bad rows; counts match both timesA screenshot of the pipeline running without errorsA larger month of clean data that loads just as quicklyA note in the README saying the pipeline was tested by hand

Why each option is right or wrong

“A screenshot of the pipeline running without errors” (Incorrect): Running without errors says nothing about whether the numbers are right.
“A larger month of clean data that loads just as quickly” (Incorrect): Clean data doesn't test the hard cases.
“A note in the README saying the pipeline was tested by hand” (Incorrect): A note isn't evidence others can repeat.
“Run it twice on a file with bad rows; counts match both times” is correct.
A re-run that changes nothing, and counts that add up with bad rows present, show the load is safe to repeat and accounts for each row.
- Show: rows in = loaded + quarantined + removed.
- Also: what happens when a file is missing.

📖 Revise: Lesson 2: Project 1: a pipeline you can trust

Question ID path-test-data-q39 · Report a mistake in this question

Question 40 of 40 · DT10Your capstone case study says "this pipeline cut SabziSathi's costs by 40%". What's wrong, and how should it be written?Nothing; case studies are expected to sound impressive to employersUse measured kit figures, label illustrations, say what's untestedChange it to 25%, which sounds more believable to readersRemove all numbers, since they can't be checked anyway

Why each option is right or wrong

“Nothing; case studies are expected to sound impressive to employers” (Incorrect): Claims that can't be backed up damage trust when someone asks.
“Change it to 25%, which sounds more believable to readers” (Incorrect): A smaller invented number is still invented.
“Remove all numbers, since they can't be checked anyway” (Incorrect): Measured numbers from the kit can be checked and make the case study stronger.
“Use measured kit figures, label illustrations, say what's untested” is correct.
SabziSathi is a practice kit, so there are no real costs. Measured facts (tests passed, rows reconciled, data scanned) and clear labels are honest and still convincing.
- Write: "illustrative" for any estimate.
- Add: what you'd test next in a real setting.

📖 Revise: Lesson 5: Demo, self-review and case studies

Question ID path-test-data-q40 · Report a mistake in this question