Job search
SQL Interview Practice Platforms: Compare Database Engines, Hints, and Paid Access
A locally executed calendar-window exercise and a completed engine-first buying decision separate working SQL, useful hints, and paid feedback.
Choose a SQL interview practice platform by the database you must use, the help you need, and the kind of feedback your next attempt requires. A large question count cannot compensate for practicing the wrong date syntax or reading solutions without learning to explain the result. Attempt one problem in the required dialect and record where you need help before paying for broader access.
We inspected DataLemur, StrataScratch, and SQLPad public offers on October 20, 2026. We opened a public hint and checked engine menus and pricing; we did not buy access, sign in, submit answers, or benchmark their graders. The worked SQL below was executed locally in SQLite, with other dialects discussed from official documentation. Tie the exercise to a target role using recent job listings on LandOffer, then confirm the actual interview engine with the recruiter.

Match the database before comparing question libraries
DataLemur's inspected SQL question displayed PostgreSQL 14 above the editor, alongside Run Code and Submit. That is a useful observed starting point for PostgreSQL preparation. It does not establish the runtime version of every question or prove that an employer uses that version. We did not execute its starter query or test alternative engines.
StrataScratch's coding catalog had an expanded menu containing Postgres, MySQL, MSSQL, and Oracle, as well as Pandas, Polars, PySpark, and R. The four SQL labels make it worth investigating when a recruiter specifies a dialect beyond PostgreSQL. A filter label still does not prove that every selected problem has equivalent solutions or identical execution behavior in each database. Inspect the particular task you plan to practice.
SQLPad's pricing page describes Postgres and MySQL practice in its SQL & R tier. Its FAQ also lists engine versions, but that section is dated March 21, 2023. Treat those versions as an old documentation statement rather than a current runtime probe. SQLPad.io, the interview preparation service reviewed here, should also be distinguished from the unrelated open-source SQLPad database client.
Basic joins and aggregation transfer across engines, but date arithmetic, timestamp conversion, identifiers, and window-frame support can differ. You can understand a metric and still lose time recalling the right date function. Conversely, a syntactically valid query may answer a different question. The following exercise isolates those two problems without copying a vendor's interview question.
Worked SQL: three calendar days are different from three rows
This original synthetic table records tickets opened on three dates. Assume each row is already aggregated to one calendar day, dates use a consistent business timezone, and an absent date means zero tickets rather than missing ingestion. Return a trailing three-calendar-day total for each date present in the table, including the current day.
| Day | Tickets opened | Expected three-day total |
|---|---|---|
| October 1, 2026 | 5 | 5 |
| October 3, 2026 | 7 | 12 |
| October 4, 2026 | 9 | 16 |
For October 4, the requested interval is October 2 through October 4. The five tickets on October 1 are outside it. A window using ROWS BETWEEN 2 PRECEDING AND CURRENT ROW instead takes the three available rows and returns 21. It executes successfully while answering “last three recorded rows.” Adding a date to the sort does not change that frame's meaning.
Here is the complete SQLite query used for our local check, with tickets(day, opened) containing the rows above. ISO date strings are deliberate; this is not a query over unnormalized timestamps.
SELECT a.day,
(SELECT SUM(b.opened)
FROM tickets AS b
WHERE b.day BETWEEN date(a.day, '-2 days')
AND a.day) AS three_day_total
FROM tickets AS a
ORDER BY a.day;
The executed result was 5, 12, and 16 in date order. The row-frame alternative returned 5, 12, and 21. We then inserted an explicit October 2 row with zero tickets: October 4 remained 16. Changing October 1 from 5 to 500 also left October 4 at 16, confirming that an outside-boundary value did not enter that day's interval. These are deterministic checks of our own fixture, not DataLemur, StrataScratch, or SQLPad scores.
The query only returns dates present in the table. A report that must display every calendar date needs a calendar spine or equivalent generated date series; otherwise October 2 will never appear, even though its assumed count is zero. If absent rows represent a broken data feed, treating them as zero is an analytical error. State that assumption before choosing the SQL syntax.
Translate the date operation without changing the question
Keep the inclusive lower and upper bounds, and adapt the lower-bound expression to the actual type and engine. For a daily DATE column, these expressions illustrate subtracting two days:
| Engine | Lower-bound expression | Verification boundary |
|---|---|---|
| SQLite | date(a.day, '-2 days') |
Executed locally on the synthetic table. |
| PostgreSQL | a.day - 2 |
Documented date-minus-integer operation; assumes day is a DATE. |
| MySQL | DATE_SUB(a.day, INTERVAL 2 DAY) |
Official date arithmetic syntax; not executed here. |
| SQL Server | DATEADD(day, -2, a.day) |
Official date arithmetic syntax; not executed here. |
PostgreSQL's date documentation, MySQL's date functions, and SQL Server's DATEADD reference support those translations. They are not evidence that a practice platform runs the same current release. The official MySQL page loaded natively after a command-line fetch was denied.
We used a correlated sum to make the boundary visible, not to claim it is the fastest production implementation. A date-range window or a dense daily series may offer another solution depending on the engine and task. In an interview, explain the interval and missing-day rule first, then discuss performance with the actual table size and indexing. Check column types before translating: the PostgreSQL expression assumes a DATE, while a timestamp input requires an explicit decision about time-of-day and day boundaries.

Separate hints, full solutions, and evidence of learning
On DataLemur, we opened the first public hint on the inspected question. Choosing Solution then requested sign-in. That observation establishes a public hint and an authentication gate on one question; it does not prove that every solution requires a paid plan. Its Premium page advertises multiple hints and full solutions. Check the access state of the exact questions you need before deciding what an upgrade supplies.
StrataScratch's current comparison matrix gives a free subset of coding questions, hints, walkthroughs, and user solutions, with full access on paid plans. The matrix separately marks AI guidance as Premium. A Free plan's AI credit balance therefore does not establish access to every AI feature. Coming-soon rehearsal features should not be counted as available feedback.
SQLPad's FAQ describes ten free SQL/Python questions before the account reaches its free limit; its SQL & R paid offer advertises coding questions with solutions and a course. We did not exhaust that allowance or inspect a paid answer. A published sample, a free account allowance, and unrestricted paid-library access are different states.
Use help to repair a named reasoning gap. For our exercise, a useful hint would direct attention to the missing October 2 row or ask what “preceding” counts. A full answer may reveal the date expression, but seeing it does not establish that you can reconstruct the interval. After reviewing help, close it and change October 1 to 500 tickets while keeping October 3 and 4 fixed. Predict the result before running again. Our 500-ticket variation makes a memorized row-frame answer visibly fail.
An accepted query is useful evidence about the judge's tests. It does not by itself measure your explanation, performance tradeoffs, or the completeness of your assumptions. A reviewer should ask why October 1 is excluded and whether absent days mean zero. Human feedback is most useful when you need that dialogue; AI guidance and static explanations can help, but their advertised availability is not a measured accuracy claim.
Compare the charge with the access you actually need
DataLemur displayed $15 monthly and $60 yearly. The $5 monthly equivalent on the yearly offer describes an annual purchase, not a $5 one-month subscription. Its $300 one-time coaching, book, and lifetime-access bundle is a separate scope. Do not buy the bundle just to unlock SQL questions if the extra deliverables do not address your gap.
StrataScratch displayed $24 monthly, $139 annually, and $289 lifetime before applying any promotional code. Monthly and yearly offers list 1,000 AI credits each month; lifetime lists 25,000 one-time nonexpiring credits. Those balances are separate from full question access and do not establish unlimited AI feedback or a fixed number of mocks.
SQLPad's page initially showed quarterly billing selected, with SQL & R at $79 per month, Python at $99, and AI at $149. The public calculator represented the SQL quarter as $237. Two attempts to switch the public billing control left the displayed state unchanged during our inspection, so we cannot report a successfully observed one-month purchase offer. Its separate SQL & R annual pass was $599 for twelve months, advertised as one payment without recurring charges. Confirm the charged period rather than treating $79 as the entire order.
These are dated public offers, not checkout totals. Renewal, cancellation, refund eligibility, taxes, and currency need purchase-specific confirmation. StrataScratch's page advertises a five-day guarantee, while its general terms, dated November 11, 2020 in our earlier review, state a general nonrefund position. Confirm which purchase-specific rule applies; this dated conflict does not establish an exception. Our comparison makes no refund-test claim. Avoid upgrading to AI or Python tiers merely because their additional features share the pricing page.
A completed platform and feedback decision
Arun is a fictional analytics engineer candidate with a SQL Server screen next week. He can write joins but confused row counts with calendar intervals on the exercise. His immediate deliverables are a correct query, an explanation of missing dates, and a fresh attempt after feedback. He does not need Python projects, a certificate, or an AI mock bundle.
| Candidate check | Completed decision | Reason and unresolved boundary |
|---|---|---|
| Engine | Investigate StrataScratch's MSSQL-filtered free subset first. | MSSQL is an observed catalog option; availability and help for the chosen date-window problem still need checking. |
| Concept repair | Use the local fixture to explain 16 versus 21, then translate the lower bound with DATEADD. | This isolates the frame error; local SQLite success does not establish SQL Server execution. |
| Paid access | Consider one $24 month only if the relevant exercise or explanation is outside the free subset. | One screen does not establish a reason for an annual or lifetime commitment. |
| Other offers | Keep DataLemur's PostgreSQL hints as supplementary reasoning practice; defer SQLPad purchase. | Neither inspected offer establishes an MSSQL environment; SQLPad's quarterly charge also exceeds a one-month headline comparison. |
| Feedback | Ask a study partner to challenge the zero-day assumption and request a date-boundary variation. | No automated score was observed to assess this spoken explanation. |
Arun's acceptance rule is concrete: explain the expected October 4 result before coding, produce 16 after the outside-boundary mutation, and state when a date spine is necessary. If he cannot do that, return to the interval definition before adding more questions to the queue. The failed explanation gives him a specific next step. If the actual interview permits any SQL dialect, his engine constraint changes and the PostgreSQL route becomes more plausible; ask before buying around an assumption.
Use a short independent attempt to choose the platform, then buy only missing access or feedback. Keep the date, engine, expected output, hint used, and next independent result in your practice notes. That record is more useful than a growing solved counter without an explanation. Browse recent LandOffer roles with the actual preparation requirement in mind.



