Marius from D8A

Founder of D8A
Helping data analyst getting their first job

1K+ data folks on my discord
Follow for one cheatsheet every day

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/z7aWmj4e6c

vaguest requirement you ever received? 👇

[Marius from D8A @MariusFromD8A]

Stop asking stakeholders what they want.

One notebook page: why the wish list question fails, who has to be in the room, the four questions that surface the real problem, how to capture it, and how to end with a ranked list and an owner.

[Marius from D8A @MariusFromD8A]

The recap that makes it real: Seven lines, sent within 24 hours. Fill the brackets while the room still remembers.

Trap: the requirement is a solution: Three things you will hear in every workshop, and the question that opens them up.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/PnC3e4iSV7

most expensive query you ever ran? 👇

[Marius from D8A @MariusFromD8A]

Your query scans 2 TB to read one day.

One notebook page: what a partition is, how pruning skips the days you do not need, the filters that silently turn it off, clustering as the second lever, and the BigQuery guardrails.

[Marius from D8A @MariusFromD8A]

Coarse partition, fine cluster: Partition on the date everyone filters by. Cluster on the columns they filter next.

Trap: the filter is there, the scan is full: Three BigQuery filters that look selective and read every partition anyway.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/6O39RaBA7G

what is the worst duplicate bug you ever shipped? 👇

[Marius from D8A @MariusFromD8A]

DISTINCT does not dedupe your table.

One notebook page: what a duplicate really is, how to find them, the ROW_NUMBER pattern that keeps the latest row, the tie-breaker rule, and how to stop them at the source.

[Marius from D8A @MariusFromD8A]

Keep the latest row per key: Number the rows inside each key, newest first, and keep number one.

Trap: dedupes that do not dedupe: Each one passes a quick look. None of them gives one row per key, every run.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/beAqQo8KXj

which SQL interview question caught you off guard? 👇

[Marius from D8A @MariusFromD8A]

LIMIT 1 OFFSET 1 fails the interview.

One notebook page: the second highest salary question, why the naive answer breaks on ties and one-row tables, the three robust answers, and the nth and per-department variants.

[Marius from D8A @MariusFromD8A]

The answer that survives a tie: Rank the salary values, not the rows. Ties share a rank, and an empty result becomes NULL.

Trap: same question, three tables: The naive query is right on clean data. Interview data is never clean.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/x7uJgftMbp

which of these need scaling: KNN, random forest, logistic regression, XGBoost? 👇

[Marius from D8A @MariusFromD8A]

You scaled your data for a random forest.

One notebook page: the two formulas, which models need scaling and which ignore it, which scaler to pick, and the leakage mistake that makes your CV score a lie.

[Marius from D8A @MariusFromD8A]

The scaler lives inside the Pipeline: Fit on train, transform test, never the other way round. The Pipeline makes that impossible to get wrong.

Trap: three scaling mistakes: Two are wasted work, one is a leak that inflates your score.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/E2hKLeSMmy

what broke your pipeline last? column rename, type change, late file? 👇

[Marius from D8A @MariusFromD8A]

Upstream renamed a column. You found out at 3am.

One notebook page: what a data contract is, what it must cover, where it lives so it is enforced, which changes break it, and how to roll a breaking change out without a pager.

[Marius from D8A @MariusFromD8A]

A contract is a file, not a meeting: Twelve lines next to the producer code. CI reads it, consumers read it, nobody is surprised.

Trap: the same change, with and without: Three real incidents and how a contract turns each one into a blocked pull request.

[Marius from D8A @MariusFromD8A]

Full sheet as a free print-ready PDF, no signup:
https://t.co/OqPyps0zeg

type 1 or type 2 for a customer email change? defend it 👇

[Marius from D8A @MariusFromD8A]

Your customer moved. Your warehouse disagrees.

One notebook page: why dimensions need versions, the three SCD types with what each one loses, the surrogate-key rule, and the join that silently doubles revenue.

[Marius from D8A @MariusFromD8A]

Type 2, in two statements: Close the current version, open a new one. Facts written from now on join the new surrogate key.

Trap: the report that changed overnight: Three SCD mistakes and how each one shows up in a dashboard.