Most Candidates Fail These SQL Ideas in Information Interviews
Picture by writer | Canva
An interviewer’s job is to search out essentially the most appropriate candidates for the marketed place. In doing so, they may gladly arrange SQL interview inquiries to see if they will catch you off guard. There are a number of SQL ideas at which candidates usually fail.
Hopefully, you’ll be a kind of who keep away from that future, as I’ll clarify these ideas intimately under, full with examples of remedy sure issues accurately.
# 1. Window Features
Why It’s Laborious: Candidates memorize what every window function does however don’t actually perceive how window frames, partitions, or ordering really work.
Widespread Errors: A typical mistake just isn’t specifying ORDER BY in rating window capabilities or worth window capabilities, akin to LEAD() or LAG(), and anticipating the question to work or for the outcome to be deterministic.
Instance: In this example, it is advisable to discover customers who made a second buy inside 7 days of any earlier buy.
You would possibly write this question.
WITH ordered_tx AS (
SELECT user_id,
created_at::date AS tx_date,
LAG(created_at::DATE) OVER (PARTITION BY user_id) AS prev_tx_date
FROM amazon_transactions
)
SELECT DISTINCT user_id
FROM ordered_tx
WHERE prev_tx_date IS NOT NULL AND tx_date - prev_tx_date <= 7;
At first look, all the things might sound proper. The code even outputs one thing that may seem like an accurate reply.
Initially, we’re fortunate that the code works in any respect! This occurs just because I’m writing it in PostgreSQL. In another SQL flavors, you’d get an error since ORDER BY is necessary in rating and analytical window capabilities.
Second, the output is flawed; I highlighted some rows that shouldn’t be there. Why do they seem, then?
They seem as a result of we didn’t specify an ORDER BY clause within the LAG() window operate. With out it, the row order is unfair. So, we’re evaluating the present transaction to some random earlier row for that consumer, not the one which occurred instantly earlier than it in time.
This isn’t what the query asks. We have to examine every transaction to the earlier one by date. In different phrases, we have to specify this explicitly within the ORDER BY clause inside the LAG() operate.
WITH ordered_tx AS (
SELECT user_id,
created_at::date AS tx_date,
LAG(created_at::DATE) OVER (PARTITION BY user_id ORDER BY created_at) AS prev_tx_date
FROM amazon_transactions
)
SELECT DISTINCT user_id
FROM ordered_tx
WHERE prev_tx_date IS NOT NULL AND tx_date - prev_tx_date <= 7;
# 2. Filtering With Aggregates (Particularly HAVING vs. WHERE)
Why It’s Laborious: Folks usually don’t perceive the execution order in SQL, which is: FROM -> WHERE -> GROUP BY -> HAVING -> SELECT -> ORDER BY. This order implies that WHERE filters rows before aggregation, and HAVING filters after. That additionally, logically, means which you can’t use mixture capabilities within the WHERE clause.
Widespread Mistake: Attempting to make use of mixture capabilities in WHERE in a grouped question and getting an error.
Instance: This interview question asks you to search out the full income made by every vineyard. Solely wineries the place 90 is the bottom variety of factors for any of their varieties needs to be thought-about.
Many will see this as a straightforward query and swiftly write this question.
SELECT vineyard,
selection,
SUM(value) AS total_revenue
FROM winemag_p1
WHERE MIN(factors) >= 90
GROUP BY vineyard, selection
ORDER BY vineyard, total_revenue DESC;
Nevertheless, that code will throw an error stating that mixture capabilities are usually not allowed within the WHERE clause. This gorgeous a lot explains all the things. The answer? Transfer the filtering situation from WHERE to HAVING.
SELECT vineyard,
selection,
SUM(value) AS total_revenue
FROM winemag_p1
GROUP BY vineyard, selection
HAVING MIN(factors) >= 90
ORDER BY vineyard, total_revenue DESC;
# 3. Self-Joins for Time-Based mostly or Occasion-Based mostly Comparisons
Why It’s Laborious: The thought of joining a table with itself is sort of unintuitive, so candidates usually overlook it’s an choice.
Widespread Mistake: Utilizing subqueries and complicating the question when becoming a member of a desk with itself can be less complicated and quicker, particularly when filtering by dates or occasions.
Instance: Right here’s a question asking you to point out the change of each forex’s change charge between 1 January 2020 and 1 July 2020.
You’ll be able to remedy this by writing an outer correlated subquery that fetches the July 1 change charges, then subtracts the January 1 change charges, which come from the internal subquery.
SELECT jan_rates.source_currency,
(SELECT exchange_rate
FROM sf_exchange_rate
WHERE source_currency = jan_rates.source_currency AND date="2020-07-01") - jan_rates.exchange_rate AS distinction
FROM (SELECT source_currency, exchange_rate
FROM sf_exchange_rate
WHERE date="2020-01-01"
) AS jan_rates;
This returns an accurate output, however such an answer is unnecessarily sophisticated. A a lot less complicated resolution, with fewer strains of code, includes self-joining a desk with itself after which making use of two date filtering circumstances within the WHERE clause.
SELECT jan.source_currency,
jul.exchange_rate - jan.exchange_rate AS distinction
FROM sf_exchange_rate jan
JOIN sf_exchange_rate jul ON jan.source_currency = jul.source_currency
WHERE jan.date="2020-01-01" AND jul.date="2020-07-01";
# 4. Subqueries vs. Widespread Desk Expressions (CTEs)
Why It’s Laborious: Folks usually get caught on subqueries as a result of they study them earlier than Widespread Desk Expressions (CTEs) and proceed utilizing them for any question with layered logic. Nevertheless, subqueries can get messy in a short time.
Widespread Mistake: Utilizing deeply nested SELECT statements when CTEs can be a lot less complicated.
Instance: Within the interview question from Google and Netflix, it is advisable to discover the highest actors primarily based on their common film score inside the style through which they seem most regularly.
The answer utilizing CTEs is as follows.
WITH genre_stats AS
(SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name,
style),
max_genre_count AS
(SELECT actor_name,
MAX(movie_count) AS max_count
FROM genre_stats
GROUP BY actor_name),
top_genres AS
(SELECT gs.*
FROM genre_stats gs
JOIN max_genre_count mgc ON gs.actor_name = mgc.actor_name
AND gs.movie_count = mgc.max_count),
top_genre_avg AS
(SELECT actor_name,
MAX(avg_rating) AS max_avg_rating
FROM top_genres
GROUP BY actor_name),
filtered_top_genres AS
(SELECT tg.*
FROM top_genres tg
JOIN top_genre_avg tga ON tg.actor_name = tga.actor_name
AND tg.avg_rating = tga.max_avg_rating),
ranked_actors AS
(SELECT *,
DENSE_RANK() OVER (
ORDER BY avg_rating DESC) AS rank
FROM filtered_top_genres),
final_selection AS
(SELECT MAX(rank) AS max_rank
FROM ranked_actors
WHERE rank <= 3)
SELECT actor_name,
style,
avg_rating
FROM ranked_actors
WHERE rank <=
(SELECT max_rank
FROM final_selection);
It’s comparatively sophisticated, but it surely nonetheless consists of six clear CTEs, with the code’s readability enhanced by clear aliases.
Curious what the identical resolution would appear to be utilizing solely subqueries? Right here it’s.
SELECT ra.actor_name,
ra.style,
ra.avg_rating
FROM (
SELECT *,
DENSE_RANK() OVER (ORDER BY avg_rating DESC) AS rank
FROM (
SELECT tg.*
FROM (
SELECT gs.*
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS gs
JOIN (
SELECT actor_name,
MAX(movie_count) AS max_count
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS genre_stats
GROUP BY actor_name
) AS mgc
ON gs.actor_name = mgc.actor_name AND gs.movie_count = mgc.max_count
) AS tg
JOIN (
SELECT actor_name,
MAX(avg_rating) AS max_avg_rating
FROM (
SELECT gs.*
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS gs
JOIN (
SELECT actor_name,
MAX(movie_count) AS max_count
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS genre_stats
GROUP BY actor_name
) AS mgc
ON gs.actor_name = mgc.actor_name AND gs.movie_count = mgc.max_count
) AS top_genres
GROUP BY actor_name
) AS tga
ON tg.actor_name = tga.actor_name AND tg.avg_rating = tga.max_avg_rating
) AS filtered_top_genres
) AS ra
WHERE ra.rank <= (
SELECT MAX(rank)
FROM (
SELECT *,
DENSE_RANK() OVER (ORDER BY avg_rating DESC) AS rank
FROM (
SELECT tg.*
FROM (
SELECT gs.*
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS gs
JOIN (
SELECT actor_name,
MAX(movie_count) AS max_count
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS genre_stats
GROUP BY actor_name
) AS mgc
ON gs.actor_name = mgc.actor_name AND gs.movie_count = mgc.max_count
) AS tg
JOIN (
SELECT actor_name,
MAX(avg_rating) AS max_avg_rating
FROM (
SELECT gs.*
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS gs
JOIN (
SELECT actor_name,
MAX(movie_count) AS max_count
FROM (
SELECT actor_name,
style,
COUNT(*) AS movie_count,
AVG(movie_rating) AS avg_rating
FROM top_actors_rating
GROUP BY actor_name, style
) AS genre_stats
GROUP BY actor_name
) AS mgc
ON gs.actor_name = mgc.actor_name AND gs.movie_count = mgc.max_count
) AS top_genres
GROUP BY actor_name
) AS tga
ON tg.actor_name = tga.actor_name AND tg.avg_rating = tga.max_avg_rating
) AS filtered_top_genres
) AS ranked_actors
WHERE rank <= 3
);
There may be redundant logic repeated throughout subqueries. What number of subqueries is that? I don’t know. The code is unattainable to keep up. Although I simply wrote it, I’d nonetheless want half a day to know it if I wished to vary one thing tomorrow. Moreover, the fully meaningless subquery aliases don’t assist.
# 5. Dealing with NULLs in Logic
Why It’s Laborious: Candidates usually suppose that NULL is the same as one thing. It’s not. NULL isn’t equal to something — not even itself. Logic involving NULLs behaves in a different way from logic involving precise values.
Widespread Mistake: Utilizing = NULL as a substitute of IS NULL in filtering or lacking output rows as a result of NULLs break the situation logic.
Instance: There’s an interview question by IBM that asks you to calculate the full variety of interactions and the full variety of contents created for every buyer.
It doesn’t sound too tough, so that you would possibly write this resolution with two CTEs, the place one CTE counts the variety of interactions per buyer, whereas the opposite counts the variety of content material gadgets created by a buyer. Within the remaining SELECT, you FULL OUTER JOIN the 2 CTEs, and you’ve got the answer. Proper?
WITH interactions_summary AS
(SELECT customer_id,
COUNT(*) AS total_interactions
FROM customer_interactions
GROUP BY customer_id),
content_summary AS
(SELECT customer_id,
COUNT(*) AS total_content_items
FROM user_content
GROUP BY customer_id)
SELECT i.customer_id,
i.total_interactions,
c.total_content_items
FROM interactions_summary AS i
FULL OUTER JOIN content_summary AS c ON i.customer_id = c.customer_id
ORDER BY customer_id;
Virtually proper. Right here’s the output. (By the best way, you see double citation marks (“”) as a substitute of NULL. That’s how the StrataScratch UI shows it, however belief me, the engine nonetheless treats them for what they’re: NULL values).
The highlighted rows comprise NULLs. This makes the output incorrect. A NULL worth is neither the client ID nor the variety of interactions and contents, which the query explicitly asks you to point out.
What we’re lacking within the above resolution is COALESCE() to deal with NULLs within the remaining SELECT. Now, all the shoppers with out interactions will get their IDs from the content_summary CTE. Additionally, for purchasers that don’t have interactions, or content material, or each, we’ll now exchange NULL with 0, which is a legitimate quantity.
WITH interactions_summary AS
(SELECT customer_id,
COUNT(*) AS total_interactions
FROM customer_interactions
GROUP BY customer_id),
content_summary AS
(SELECT customer_id,
COUNT(*) AS total_content_items
FROM user_content
GROUP BY customer_id)
SELECT COALESCE(i.customer_id, c.customer_id) AS customer_id,
COALESCE(i.total_interactions, 0) AS total_interactions,
COALESCE(c.total_content_items, 0) AS total_content_items
FROM interactions_summary AS i
FULL OUTER JOIN content_summary AS c ON i.customer_id = c.customer_id
ORDER BY customer_id;
# 6. Group-Based mostly Deduplication
Why It’s Laborious: Group-based deduplication means you’re deciding on one row per group, e.g., “most up-to-date”, “highest rating”, and many others. At first, it sounds such as you solely want to choose one row per consumer. However you’ll be able to’t use GROUP BY except you mixture. However, you usually want a full row, not a single worth that aggregation and GROUP BY return.
Widespread Mistake: Utilizing GROUP BY + LIMIT 1 (or DISTINCT ON, which is PostgreSQL-specific) as a substitute of ROW_NUMBER() or RANK(), the latter if you would like ties included.
Instance: This question asks you to determine the best-selling merchandise for every month, and there’s no must separate months by 12 months. The very best-selling merchandise is calculated as unitprice * amount.
The naive method can be this. First, extract the sale month from invoicedate, choose description, and discover the full gross sales by summing unitprice * amount. Then, to get the full gross sales by month and product description, we merely GROUP BY these two columns. Lastly, we solely want to make use of ORDER BY to kind the output from the most effective to the worst-selling product and use LIMIT 1 to output solely the primary row, i.e., the best-selling merchandise.
SELECT DATE_PART('MONTH', invoicedate) AS sale_month,
description,
SUM(unitprice * amount) AS total_paid
FROM online_retail
GROUP BY sale_month, description
ORDER BY total_paid DESC
LIMIT 1;
As I stated, that is naive; the output considerably resembles what we want, however we want this for each month, not only one.
One of many right approaches is to make use of the RANK() window operate. With this method, we comply with an analogous technique to the earlier code. The distinction is that the question now turns into a subquery within the FROM clause. As well as, we use RANK() to partition the info by month after which rank the rows inside every partition (i.e., for every month individually) from the best-selling to the worst-selling merchandise.
Then, in the primary question, we merely choose the required columns and output solely rows the place the rank is 1 utilizing the WHERE clause.
SELECT month,
description,
total_paid
FROM
(SELECT DATE_PART('month', invoicedate) AS month,
description,
SUM(unitprice * amount) AS total_paid,
RANK() OVER (PARTITION BY DATE_PART('month', invoicedate) ORDER BY SUM(unitprice * amount) DESC) AS rnk
FROM online_retail
GROUP BY month, description) AS tmp
WHERE rnk = 1;
# Conclusion
The six ideas we’ve lined generally seem in SQL coding interview questions. Take note of them, follow interview questions that contain these ideas, study the right approaches, and also you’ll considerably enhance your probabilities in your interviews.
Nate Rosidi is an information scientist and in product technique. He is additionally an adjunct professor instructing analytics, and is the founding father of StrataScratch, a platform serving to information scientists put together for his or her interviews with actual interview questions from prime firms. Nate writes on the newest developments within the profession market, provides interview recommendation, shares information science tasks, and covers all the things SQL.
